2026
|
George Charalambous, George Moustris, Costas Tzafestas Beyond Task-Specific Models: A Multi-Task Transformer for Robotic Surgery Skill Assessment Journal Article In: IEEE Transactions on Medical Robotics and Bionics, 2026. @article{nokey,
title = {Beyond Task-Specific Models: A Multi-Task Transformer for Robotic Surgery Skill Assessment},
author = {George Charalambous, George Moustris, Costas Tzafestas},
url = {https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11744486},
doi = {10.1109/TMRB.2026.3745199},
year = {2026},
date = {2026-10-21},
urldate = {2026-10-21},
journal = {IEEE Transactions on Medical Robotics and Bionics},
publisher = {IEEE Transaction on Medical Robotics & Bionics},
organization = {IEEE},
abstract = {Automatic assessment of surgical skills is a critical component in ensuring valid and high-quality surgical training, with respect to patient safety and clinical care standards. In this paper, we propose a Multi-Task Transformer model that utilizes kinematic data collected from a surgical robot to evaluate surgical skills. We employ the widely used JIGSAWS dataset for training and evaluation of the model, which comprises structured and annotated recordings of surgical tasks from the Da Vinci surgical system. Our proposed model produces an estimate of all six domains of the Modified Objective Structured Assessment of Technical Skills (MOSATS) scale, as well as the overall Global Rating Score (GRS). The experimental results demonstrate a strong correlation between the model’s predictions and expert human evaluations, outperforming state-of-the-art models that rely on kinematic data in 83% of domain-task combinations, achieving a mean Spearman correlation coefficient (SCC) of 0.699 (vs. 0.59). Additionally, the performance of our model in predicting GRS scores (SCC: 0.84) matches or exceeds several vision-based approaches for individual tasks. The model’s capacity for generalization across a range of surgical tasks demonstrates its potential as a reliable tool for providing feedback and assistance in surgical training. Source code and models are available at https://github.com/gchar00/MultiTask-Transformer.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Automatic assessment of surgical skills is a critical component in ensuring valid and high-quality surgical training, with respect to patient safety and clinical care standards. In this paper, we propose a Multi-Task Transformer model that utilizes kinematic data collected from a surgical robot to evaluate surgical skills. We employ the widely used JIGSAWS dataset for training and evaluation of the model, which comprises structured and annotated recordings of surgical tasks from the Da Vinci surgical system. Our proposed model produces an estimate of all six domains of the Modified Objective Structured Assessment of Technical Skills (MOSATS) scale, as well as the overall Global Rating Score (GRS). The experimental results demonstrate a strong correlation between the model’s predictions and expert human evaluations, outperforming state-of-the-art models that rely on kinematic data in 83% of domain-task combinations, achieving a mean Spearman correlation coefficient (SCC) of 0.699 (vs. 0.59). Additionally, the performance of our model in predicting GRS scores (SCC: 0.84) matches or exceeds several vision-based approaches for individual tasks. The model’s capacity for generalization across a range of surgical tasks demonstrates its potential as a reliable tool for providing feedback and assistance in surgical training. Source code and models are available at https://github.com/gchar00/MultiTask-Transformer. |
Konstantinos Fotopoulos, Petros Maragos, Panagiotis Misiakos TropNNC: Structured Neural Network Compression Using Tropical Geometry Journal Article In: Transactions on Machine Learning Research, 2026. @article{nokey,
title = {TropNNC: Structured Neural Network Compression Using Tropical Geometry},
author = {Konstantinos Fotopoulos, Petros Maragos, Panagiotis Misiakos},
url = {https://openreview.net/pdf?id=u7DRq1icmY},
year = {2026},
date = {2026-06-01},
urldate = {2026-06-01},
journal = {Transactions on Machine Learning Research},
abstract = {We present TropNNC, a framework for compressing neural networks with linear and convolutional layers and ReLU-type activations using tropical geometry. By representing a network’s output as a tropical rational function, TropNNC enables structured compression via reduction of the corresponding tropical polynomials. Our method identifies redundancy via similarity and improves upon the geometric approximation of previous work by adaptively selecting the weights of retained neurons. We relate it to SVD and spectral clustering, and provide insights into network compression beyond the specific setting considered. We provide the tightest known theoretical compression bound, and the first successful application of tropical geometry to convolutional layers. TropNNC requires access only to network weights – no training data – and achieves competitive performance on MNIST, CIFAR, and ImageNet, matching strong baselines such as ThiNet and CUP.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
We present TropNNC, a framework for compressing neural networks with linear and convolutional layers and ReLU-type activations using tropical geometry. By representing a network’s output as a tropical rational function, TropNNC enables structured compression via reduction of the corresponding tropical polynomials. Our method identifies redundancy via similarity and improves upon the geometric approximation of previous work by adaptively selecting the weights of retained neurons. We relate it to SVD and spectral clustering, and provide insights into network compression beyond the specific setting considered. We provide the tightest known theoretical compression bound, and the first successful application of tropical geometry to convolutional layers. TropNNC requires access only to network weights – no training data – and achieves competitive performance on MNIST, CIFAR, and ImageNet, matching strong baselines such as ThiNet and CUP. |
Siyuan Li, Shikai Fang, Lei Cheng, Yik-Chung Wu, Sergios Theodoridis When VMP Meets CEP: An Algorithmic Equivalence Under Mild Conditions Journal Article In: Transactions on Machine Learning Research, 2026. @article{nokey,
title = {When VMP Meets CEP: An Algorithmic Equivalence Under Mild Conditions},
author = {Siyuan Li, Shikai Fang, Lei Cheng, Yik-Chung Wu, Sergios Theodoridis},
url = {https://openreview.net/pdf/9c3472982fe7da56816cefc1758550b06d78f218.pdf},
year = {2026},
date = {2026-05-01},
urldate = {2026-05-01},
journal = {Transactions on Machine Learning Research},
publisher = {Transactions on Machine Learning Research (TMLR)},
abstract = {Approximate Bayesian inference (ABI) methods have become indispensable tools in modern machine learning and statistics for approximating intractable posterior distributions. Despite extensive studies, the theoretical connections among different ABI methods have remained relatively unexplored. This paper establishes an algorithmic equivalence between two widely employed ABI techniques, namely variational message passing (VMP) and conditional expectation propagation (CEP). Through rigorous mathematical analysis, we demonstrate that these two approaches, despite originating from different perspectives (variational inference and expectation propagation, respectively), yield the same update equations under mild conditions, from both optimization and graphical model viewpoints. As a direct consequence, we establish a convergence guarantee for CEP and show that VMP-derived algorithms can inherit streaming variants without additional derivation effort. To validate our theoretical findings, we apply both VMP and CEP to Bayesian tensor decomposition and verify that they produce identical updates, demonstrating how the equivalence provides a principled route to a streaming variant.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Approximate Bayesian inference (ABI) methods have become indispensable tools in modern machine learning and statistics for approximating intractable posterior distributions. Despite extensive studies, the theoretical connections among different ABI methods have remained relatively unexplored. This paper establishes an algorithmic equivalence between two widely employed ABI techniques, namely variational message passing (VMP) and conditional expectation propagation (CEP). Through rigorous mathematical analysis, we demonstrate that these two approaches, despite originating from different perspectives (variational inference and expectation propagation, respectively), yield the same update equations under mild conditions, from both optimization and graphical model viewpoints. As a direct consequence, we establish a convergence guarantee for CEP and show that VMP-derived algorithms can inherit streaming variants without additional derivation effort. To validate our theoretical findings, we apply both VMP and CEP to Bayesian tensor decomposition and verify that they produce identical updates, demonstrating how the equivalence provides a principled route to a streaming variant. |
Christos Garoufis, Athanasia Zlatintsi, Petros Maragos CVSM: Contrastive Vocal Similarity Modeling Journal Article In: IEEE Transactions on Audio, Speech and Language Processing, vol. 34, pp. 2724-2737, 2026. @article{nokey,
title = {CVSM: Contrastive Vocal Similarity Modeling},
author = {Christos Garoufis, Athanasia Zlatintsi, Petros Maragos},
url = {https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11457443},
doi = {10.1109/TASLPRO.2026.3677618},
year = {2026},
date = {2026-03-01},
urldate = {2026-03-01},
journal = {IEEE Transactions on Audio, Speech and Language Processing},
volume = {34},
pages = {2724-2737},
abstract = {The availability of large, unlabeled datasets across various domains has contributed to the development of a plethora of methods that learn representations for multiple target (downstream) tasks through self-supervised pre-training. In this work, we introduce CVSM (Contrastive Vocal Similarity Modeling), a contrastive self-supervised procedure for music signal representation learning in the audio domain that can be utilized for musical and vocal similarity modeling. Our method operates under a contrastive framework, maximizing the similarity between vocal excerpts and musical mixtures containing the same vocals; we devise both a label-informed protocol, leveraging artist identity information to sample the contrastive pairs, and a label-agnostic scheme, involving artificial mixture creation from randomly sampled vocal and accompaniment excerpts, which are paired with vocals from the same audio segment. We evaluate our proposed method in measuring vocal similarity both objectively, through linear probing on a suite of appropriate downstream tasks, and subjectively, via conducting a user study consisting of pairwise comparisons between different models in a recommendation-by-query setting. Our results indicate that the representations learned through CVSM are effective in musical and vocal similarity modeling, outperforming numerous baselines across both isolated vocals and complete musical mixtures. Moreover, while the availability of artist identity labels during pre-training leads to overall more consistent performance both in the evaluated downstream tasks and the user study, a label-agnostic CVSM variant incorporating hybrid pre-training with real and artificial mixtures achieves comparable performance to the label-informed one in artist identification and perceived vocal similarity.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
The availability of large, unlabeled datasets across various domains has contributed to the development of a plethora of methods that learn representations for multiple target (downstream) tasks through self-supervised pre-training. In this work, we introduce CVSM (Contrastive Vocal Similarity Modeling), a contrastive self-supervised procedure for music signal representation learning in the audio domain that can be utilized for musical and vocal similarity modeling. Our method operates under a contrastive framework, maximizing the similarity between vocal excerpts and musical mixtures containing the same vocals; we devise both a label-informed protocol, leveraging artist identity information to sample the contrastive pairs, and a label-agnostic scheme, involving artificial mixture creation from randomly sampled vocal and accompaniment excerpts, which are paired with vocals from the same audio segment. We evaluate our proposed method in measuring vocal similarity both objectively, through linear probing on a suite of appropriate downstream tasks, and subjectively, via conducting a user study consisting of pairwise comparisons between different models in a recommendation-by-query setting. Our results indicate that the representations learned through CVSM are effective in musical and vocal similarity modeling, outperforming numerous baselines across both isolated vocals and complete musical mixtures. Moreover, while the availability of artist identity labels during pre-training leads to overall more consistent performance both in the evaluated downstream tasks and the user study, a label-agnostic CVSM variant incorporating hybrid pre-training with real and artificial mixtures achieves comparable performance to the label-informed one in artist identification and perceived vocal similarity. |
Nikolas Chatzis, Angeliki Tsinouka, Katerina Papadimitriou, Niki Efthymiou, Marios Glytsos, George Retsinas, Paris Oikonomou, Gerasimos Potamianos, Petros Maragos, Panagiotis Paraskevas Filntisis Mind the Shape Gap: A Benchmark and Baseline for Deformation-Aware 6D Pose Estimation of Agricultural Produce Conference IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026. @conference{nokey,
title = {Mind the Shape Gap: A Benchmark and Baseline for Deformation-Aware 6D Pose Estimation of Agricultural Produce},
author = {Nikolas Chatzis, Angeliki Tsinouka, Katerina Papadimitriou, Niki Efthymiou, Marios Glytsos, George Retsinas, Paris Oikonomou, Gerasimos Potamianos, Petros Maragos, Panagiotis Paraskevas Filntisis},
url = {https://arxiv.org/abs/2603.27429},
doi = {10.48550/arXiv.2603.27429},
year = {2026},
date = {2026-09-27},
urldate = {2026-09-27},
publisher = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
abstract = { Accurate 6D pose estimation for robotic harvesting is fundamentally hindered by the biological deformability and high intra-class shape variability of agricultural produce. Instance-level methods fail in this setting, as obtaining exact 3D models for every unique piece of produce is practically infeasible, while category-level approaches that rely on a fixed template suffer significant accuracy degradation when the prior deviates from the true instance geometry. To bridge such lack of robustness to deformation, we introduce PEAR (Pose and dEformation of Agricultural pRoduce), the first benchmark providing joint 6D pose and per-instance 3D deformation ground truth across 8 produce categories, acquired via a robotic manipulator for high annotation accuracy. Using PEAR, we show that state-of-the-art methods suffer up to 6× performance degradation when faced with the inherent geometric deviations of real-world produce. Motivated by this finding, we propose SEED (Simultaneous Estimation of posE and Deformation), a unified RGB-only framework that jointly predicts 6D pose and explicit lattice deformations from a single image across multiple produce categories. Trained entirely on synthetic data with generative texture augmentation applied at the UV level, SEED outperforms MegaPose on 6 out of 8 categories under identical RGB-only conditions, demonstrating that explicit shape modeling is a critical step toward reliable pose estimation in agricultural robotics.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Accurate 6D pose estimation for robotic harvesting is fundamentally hindered by the biological deformability and high intra-class shape variability of agricultural produce. Instance-level methods fail in this setting, as obtaining exact 3D models for every unique piece of produce is practically infeasible, while category-level approaches that rely on a fixed template suffer significant accuracy degradation when the prior deviates from the true instance geometry. To bridge such lack of robustness to deformation, we introduce PEAR (Pose and dEformation of Agricultural pRoduce), the first benchmark providing joint 6D pose and per-instance 3D deformation ground truth across 8 produce categories, acquired via a robotic manipulator for high annotation accuracy. Using PEAR, we show that state-of-the-art methods suffer up to 6× performance degradation when faced with the inherent geometric deviations of real-world produce. Motivated by this finding, we propose SEED (Simultaneous Estimation of posE and Deformation), a unified RGB-only framework that jointly predicts 6D pose and explicit lattice deformations from a single image across multiple produce categories. Trained entirely on synthetic data with generative texture augmentation applied at the UV level, SEED outperforms MegaPose on 6 out of 8 categories under identical RGB-only conditions, demonstrating that explicit shape modeling is a critical step toward reliable pose estimation in agricultural robotics. |
Panagiotis Mermigkas, Argyris Manetas, Petros Maragos GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition Conference IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026. @conference{nokey,
title = {GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition},
author = {Panagiotis Mermigkas, Argyris Manetas, Petros Maragos},
url = {https://arxiv.org/abs/2607.21416},
doi = {10.48550/arXiv.2607.21416},
year = {2026},
date = {2026-09-27},
urldate = {2026-09-27},
publisher = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
abstract = {Existing Gaussian-splatting-based monocular Simultaneous Localization and Mapping (SLAM) systems are either tailored to short sequences, are not real-time, or suffer from prohibitive GPU memory requirements, limiting their applicability in realistic, long-horizon scenarios. To address this, we present GLAM-SLAM, a real-time, decoupled Gaussian-splatting SLAM system designed for large-scale outdoor scenes. We ensure lightweight tracking using a robust, feature-based SLAM frontend, while for mapping, we adopt a structured, sparse anchor grid representation that ensures scalable operation and maintains scene coherence across long-term sequences. To satisfy the dense initialization requirements of 3D Gaussian Splatting (3DGS), we introduce a geometry-based flow-densification anchoring strategy using epipolar constraints. Furthermore, by treating mapping as a multi-scene problem, we propose a scene-partitioning strategy that introduces a strong spatial inductive bias via MLP initializations to generate localized Gaussians. We evaluate our system on the challenging, long-sequence KITTI Odometry, Oxford RobotCar, and Malaga datasets. Extensive ablations and comparisons demonstrate a 15% improvement in reconstruction quality over the second best performer, while maintaining real-time performance and the ability to scale to longer sequences. Code is publicly available for the benefit of the community.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Existing Gaussian-splatting-based monocular Simultaneous Localization and Mapping (SLAM) systems are either tailored to short sequences, are not real-time, or suffer from prohibitive GPU memory requirements, limiting their applicability in realistic, long-horizon scenarios. To address this, we present GLAM-SLAM, a real-time, decoupled Gaussian-splatting SLAM system designed for large-scale outdoor scenes. We ensure lightweight tracking using a robust, feature-based SLAM frontend, while for mapping, we adopt a structured, sparse anchor grid representation that ensures scalable operation and maintains scene coherence across long-term sequences. To satisfy the dense initialization requirements of 3D Gaussian Splatting (3DGS), we introduce a geometry-based flow-densification anchoring strategy using epipolar constraints. Furthermore, by treating mapping as a multi-scene problem, we propose a scene-partitioning strategy that introduces a strong spatial inductive bias via MLP initializations to generate localized Gaussians. We evaluate our system on the challenging, long-sequence KITTI Odometry, Oxford RobotCar, and Malaga datasets. Extensive ablations and comparisons demonstrate a 15% improvement in reconstruction quality over the second best performer, while maintaining real-time performance and the ability to scale to longer sequences. Code is publicly available for the benefit of the community. |
Alexandros Dimitriadis, George P. Moustris, Costas Tzafestas Deep Learning 3D Surgical Tool Pose Estimation using Geometry-Aware Correction Conference 15th Conference on New Technologies for Computer and Robot-Assisted Surgery, Leeds, UK, 2026. @conference{nokey,
title = {Deep Learning 3D Surgical Tool Pose Estimation using Geometry-Aware Correction},
author = {Alexandros Dimitriadis, George P. Moustris, Costas Tzafestas},
url = {https://www.researchgate.net/publication/407316086_Deep_Learning_3D_Surgical_Tool_Pose_Estimation_using_Geometry-Aware_Correction},
year = {2026},
date = {2026-09-16},
urldate = {2026-09-16},
publisher = {15th Conference on New Technologies for Computer and Robot-Assisted Surgery, Leeds, UK},
abstract = {In minimally invasive surgery (MIS), where visual feedback is provided through an endoscope, robust pose estimation of surgical instruments is a critical prerequisite for workflow analysis, intraoperative guidance, context-aware assistance, and automated skill assessment. While deep learning models have significantly advanced 2D instrument local- ization, extending these capabilities to robust 3D tracking via stereo endoscopy remains a challenging. Recent 2D pose architectures, such as unified multi-head networks and YOLO-based detectors, excel at extracting keypoints within individual frames. However, treating 3D reconstruction as a passive downstream calculation based on these 2D outputs is inherently fragile. Zero-shot learning architectures have been utilized to generalize pose estimation to unseen surgical instruments, but their practical deployment is hindered by their requirement for 3D CAD models of the tools and reliable depth sensing. [3]. Building upon our 2D tracking framework presented in [4], which integrates YOLOv8m and RTMPose-m with an adaptive temporal filter, this work introduces a novel, active 3D reconstruction backend. To prevent depth (𝑍- axis) explosions prior to stereo triangulation, our backend deploys self-healing mechanisms, namely cross-camera reprojection and kinematic extrapolation, to mathematically correct transient visual failures, providing an uninterrupted 3D data stream for downstream robotic control.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
In minimally invasive surgery (MIS), where visual feedback is provided through an endoscope, robust pose estimation of surgical instruments is a critical prerequisite for workflow analysis, intraoperative guidance, context-aware assistance, and automated skill assessment. While deep learning models have significantly advanced 2D instrument local- ization, extending these capabilities to robust 3D tracking via stereo endoscopy remains a challenging. Recent 2D pose architectures, such as unified multi-head networks and YOLO-based detectors, excel at extracting keypoints within individual frames. However, treating 3D reconstruction as a passive downstream calculation based on these 2D outputs is inherently fragile. Zero-shot learning architectures have been utilized to generalize pose estimation to unseen surgical instruments, but their practical deployment is hindered by their requirement for 3D CAD models of the tools and reliable depth sensing. [3]. Building upon our 2D tracking framework presented in [4], which integrates YOLOv8m and RTMPose-m with an adaptive temporal filter, this work introduces a novel, active 3D reconstruction backend. To prevent depth (𝑍- axis) explosions prior to stereo triangulation, our backend deploys self-healing mechanisms, namely cross-camera reprojection and kinematic extrapolation, to mathematically correct transient visual failures, providing an uninterrupted 3D data stream for downstream robotic control. |
Stelios Zarifis, Ioannis Kordonis, Petros Maragos Diffusion-Based Scenario Tree Generation for Multivariate Time Series Prediction and Multistage Stochastic Optimization Conference IEEE European Signal Processing Conference (EUSIPCO), 2026. @conference{nokey,
title = {Diffusion-Based Scenario Tree Generation for Multivariate Time Series Prediction and Multistage Stochastic Optimization},
author = {Stelios Zarifis, Ioannis Kordonis, Petros Maragos},
url = {https://arxiv.org/abs/2509.14832v2},
doi = {10.48550/arXiv.2509.14832},
year = {2026},
date = {2026-09-02},
urldate = {2025-09-08},
publisher = {IEEE European Signal Processing Conference (EUSIPCO)},
abstract = {Stochastic forecasting is critical for efficient decision-making in uncertain systems, such as energy markets and finance, where estimating the full distribution of future scenarios is essential. We propose Diffusion Scenario Tree (DST), a general framework for constructing scenario trees using diffusion-based probabilistic forecasting models to provide a structured model of system evolution for control tasks. DST recursively samples future trajectories and organizes them into a tree via clustering, ensuring non-anticipativity (decisions depending only on observed history) at each stage, offering a superior representation of uncertainty compared to using predictive models solely for forecasting system evolution. We integrate DST into Model Predictive Control (MPC) and evaluate it on energy arbitrage in New York State’s day-ahead electricity market. Experimental results show that our approach significantly outperforms the same optimization algorithms that use scenario trees generated by more conventional models. Furthermore, using DST for stochastic optimization yields more efficient decision policies by better handling uncertainty than deterministic and stochastic MPC variants using the same diffusion-based forecaster, and simple Model-Free Reinforcement Learning (RL) baselines.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Stochastic forecasting is critical for efficient decision-making in uncertain systems, such as energy markets and finance, where estimating the full distribution of future scenarios is essential. We propose Diffusion Scenario Tree (DST), a general framework for constructing scenario trees using diffusion-based probabilistic forecasting models to provide a structured model of system evolution for control tasks. DST recursively samples future trajectories and organizes them into a tree via clustering, ensuring non-anticipativity (decisions depending only on observed history) at each stage, offering a superior representation of uncertainty compared to using predictive models solely for forecasting system evolution. We integrate DST into Model Predictive Control (MPC) and evaluate it on energy arbitrage in New York State’s day-ahead electricity market. Experimental results show that our approach significantly outperforms the same optimization algorithms that use scenario trees generated by more conventional models. Furthermore, using DST for stochastic optimization yields more efficient decision policies by better handling uncertainty than deterministic and stochastic MPC variants using the same diffusion-based forecaster, and simple Model-Free Reinforcement Learning (RL) baselines. |
Nikolaos Tsalkitzis, Panagiotis P. Filntisis, Petros Maragos, Niki Efthymiou Uncertainty-Driven Anomaly Detection for Psychotic Relapse Using Smartwatches: Forecasting and Multi-Task Learning Fusion Conference IEEE European Signal Processing Conference (EUSIPCO), 2026. @conference{nokey,
title = {Uncertainty-Driven Anomaly Detection for Psychotic Relapse Using Smartwatches: Forecasting and Multi-Task Learning Fusion},
author = {Nikolaos Tsalkitzis, Panagiotis P. Filntisis, Petros Maragos, Niki Efthymiou},
url = {https://arxiv.org/abs/2605.13816},
doi = {10.48550/arXiv.2605.13816},
year = {2026},
date = {2026-08-31},
urldate = {2026-08-31},
publisher = {IEEE European Signal Processing Conference (EUSIPCO)},
abstract = {Digital phenotyping enables continuous passive monitoring of behavior and physiology, offering a promising paradigm for early detection of psychotic relapse. In this work, we develop and systematically study two smartwatch-based frameworks for daily relapse detection. The first forecasts cardiac dynamics and flags deviations between predicted and observed features as indicators of abnormality. The second adopts a multitask formulation that fuses sleep with motion and cardiac-derived signals, learning time-aware embeddings and predicting measurement timing. Both pipelines use Transformer encoders and output a daily anomaly score, derived from predictive uncertainty estimated via an ensemble of multilayer perceptrons to improve robustness to real-world wearable variability. While each framework independently demonstrates strong predictive power, we show that they capture complementary physiological signatures. Consequently, we propose a late-fusion strategy that synergistically combines the anomaly signals from both architectures into a unified decision score. We benchmark our methodology on the 2nd e-Prevention Grand Challenge dataset, where our fused model achieves an 8% relative improvement over the competition-winning baseline. Our results, supported by extensive ablation studies, suggest that the integration of diverse digital phenotypes, cardiac, motion, and sleep, is essential for the high-fidelity detection of psychotic relapse in real-world settings.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Digital phenotyping enables continuous passive monitoring of behavior and physiology, offering a promising paradigm for early detection of psychotic relapse. In this work, we develop and systematically study two smartwatch-based frameworks for daily relapse detection. The first forecasts cardiac dynamics and flags deviations between predicted and observed features as indicators of abnormality. The second adopts a multitask formulation that fuses sleep with motion and cardiac-derived signals, learning time-aware embeddings and predicting measurement timing. Both pipelines use Transformer encoders and output a daily anomaly score, derived from predictive uncertainty estimated via an ensemble of multilayer perceptrons to improve robustness to real-world wearable variability. While each framework independently demonstrates strong predictive power, we show that they capture complementary physiological signatures. Consequently, we propose a late-fusion strategy that synergistically combines the anomaly signals from both architectures into a unified decision score. We benchmark our methodology on the 2nd e-Prevention Grand Challenge dataset, where our fused model achieves an 8% relative improvement over the competition-winning baseline. Our results, supported by extensive ablation studies, suggest that the integration of diverse digital phenotypes, cardiac, motion, and sleep, is essential for the high-fidelity detection of psychotic relapse in real-world settings. |
Ioanna Chourdaki, Kleanthis Avramidis, Christos Garoufis, Athanasia Zlatintsi, Petros Maragos Teager-Kaiser Energy Methods for EEG Feature Extraction in Biomedical Applications Conference European Signal Processing Conference (EUSIPCO), 2026. @conference{nokey,
title = {Teager-Kaiser Energy Methods for EEG Feature Extraction in Biomedical Applications},
author = {Ioanna Chourdaki, Kleanthis Avramidis, Christos Garoufis, Athanasia Zlatintsi, Petros Maragos},
url = {https://arxiv.org/pdf/2511.17164},
doi = {10.48550/arXiv.2511.17164},
year = {2026},
date = {2026-08-31},
urldate = {2026-08-31},
publisher = {European Signal Processing Conference (EUSIPCO)},
abstract = {Electroencephalography (EEG) signals are inherently non-linear, non-stationary, and vulnerable to noise sources, making the extraction of discriminative features a long-standing challenge. In this work, we investigate the non-linear Teager-Kaiser Energy Operator (TKEO) for modeling the underlying energy dynamics of EEG in three representative tasks: motor imagery, emotion recognition, and epilepsy detection. To accommodate the narrowband nature of the operator, we employ Gabor filter banks to isolate canonical frequency bands, followed by the Energy Separation Algorithm to decompose the TKEO output into an amplitude envelope and instantaneous frequency components. We then derive a set of energy descriptors based on this demodulation and compare their classification performance against established EEG features. The proposed TKEO-based pipeline offers an intuitive, physiologically grounded framework for capturing EEG signal dynamics, while remaining simple, training-free, and data-efficient. Our findings suggest that combining TKEO features with conventional ones improves Balanced Accuracy by approximately 15% in epilepsy detection, yields modest gains in motor imagery, and achieves on par performance in emotion recognition, reflecting the pipeline’s ability to capture transient neural dynamics.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Electroencephalography (EEG) signals are inherently non-linear, non-stationary, and vulnerable to noise sources, making the extraction of discriminative features a long-standing challenge. In this work, we investigate the non-linear Teager-Kaiser Energy Operator (TKEO) for modeling the underlying energy dynamics of EEG in three representative tasks: motor imagery, emotion recognition, and epilepsy detection. To accommodate the narrowband nature of the operator, we employ Gabor filter banks to isolate canonical frequency bands, followed by the Energy Separation Algorithm to decompose the TKEO output into an amplitude envelope and instantaneous frequency components. We then derive a set of energy descriptors based on this demodulation and compare their classification performance against established EEG features. The proposed TKEO-based pipeline offers an intuitive, physiologically grounded framework for capturing EEG signal dynamics, while remaining simple, training-free, and data-efficient. Our findings suggest that combining TKEO features with conventional ones improves Balanced Accuracy by approximately 15% in epilepsy detection, yields modest gains in motor imagery, and achieves on par performance in emotion recognition, reflecting the pipeline’s ability to capture transient neural dynamics. |
Petros Georgoulas Wraight, Giorgos Sfikas, Ioannis Kordonis, Petros Maragos, George Retsinas Optimal Transport for handwritten text recognition in a low-resource regime Conference European Signal Processing Conference (EUSIPCO), 2026. @conference{nokey,
title = {Optimal Transport for handwritten text recognition in a low-resource regime},
author = {Petros Georgoulas Wraight, Giorgos Sfikas, Ioannis Kordonis, Petros Maragos, George Retsinas},
url = {https://arxiv.org/pdf/2509.16977},
doi = {10.48550/arXiv.2509.16977},
year = {2026},
date = {2026-08-31},
urldate = {2026-08-31},
publisher = {European Signal Processing Conference (EUSIPCO)},
abstract = {Handwritten Text Recognition (HTR) is a task of central importance in the field of document image understanding. State-of-the-art methods for HTR require the use of extensive annotated sets for training, making them impractical for low-resource domains like historical archives or limited-size modern collections. This paper introduces a novel framework that, unlike the standard HTR model paradigm, can leverage mild prior knowledge of lexical characteristics; this is ideal for scenarios where labeled data are scarce. We propose an iterative bootstrapping approach that aligns visual features extracted from unlabeled images with semantic word representations using Optimal Transport (OT). Starting with a minimal set of labeled examples, the framework iteratively matches word images to text labels, generates pseudo-labels for high-confidence alignments, and retrains the recognizer on the growing dataset. Numerical experiments demonstrate that our iterative visual-semantic alignment scheme significantly improves recognition accuracy on low-resource HTR benchmarks.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Handwritten Text Recognition (HTR) is a task of central importance in the field of document image understanding. State-of-the-art methods for HTR require the use of extensive annotated sets for training, making them impractical for low-resource domains like historical archives or limited-size modern collections. This paper introduces a novel framework that, unlike the standard HTR model paradigm, can leverage mild prior knowledge of lexical characteristics; this is ideal for scenarios where labeled data are scarce. We propose an iterative bootstrapping approach that aligns visual features extracted from unlabeled images with semantic word representations using Optimal Transport (OT). Starting with a minimal set of labeled examples, the framework iteratively matches word images to text labels, generates pseudo-labels for high-confidence alignments, and retrains the recognizer on the growing dataset. Numerical experiments demonstrate that our iterative visual-semantic alignment scheme significantly improves recognition accuracy on low-resource HTR benchmarks. |
Aikaterini Smyrli, Evangelos Papadopoulos Underactuated Virtual Gravity Control: Harnessing Passive Dynamics for Optimally Efficient Bipedal Locomotion Conference IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), 2026. @conference{nokey,
title = {Underactuated Virtual Gravity Control: Harnessing Passive Dynamics for Optimally Efficient Bipedal Locomotion},
author = {Aikaterini Smyrli, Evangelos Papadopoulos},
url = {https://ieeexplore.ieee.org/document/11658018},
doi = {10.1109/AIM65483.2026.11658018},
year = {2026},
date = {2026-07-07},
urldate = {2026-07-07},
publisher = {IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM)},
abstract = {This paper explores the Underactuated Virtual Gravity (UVG) controller, a model-based control approach designed to achieve energy-efficient bipedal walking. The UVG controller minimizes actuator effort during level-ground walking by capitalizing on the inherent dynamics that facilitate stable passive gaits on downward slopes. By effectively leveraging torso dynamics to support the application of the UVG, the method turns the innate underactuation of bipedal systems into an advantage. Extensive benchmarking against state-of-the-art methods such as the Trajectory Optimization for trajectory planning combined with Non-linear Model Predictive Control for tracking shows that the UVG achieves superior energy efficiency within its effective range while constituting a closed-form controller that requires fewer computational resources than numerical methods. The results highlight the UVG and related dynamics-based approaches as a compelling option for energy-conscious, task-focused robot designs.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
This paper explores the Underactuated Virtual Gravity (UVG) controller, a model-based control approach designed to achieve energy-efficient bipedal walking. The UVG controller minimizes actuator effort during level-ground walking by capitalizing on the inherent dynamics that facilitate stable passive gaits on downward slopes. By effectively leveraging torso dynamics to support the application of the UVG, the method turns the innate underactuation of bipedal systems into an advantage. Extensive benchmarking against state-of-the-art methods such as the Trajectory Optimization for trajectory planning combined with Non-linear Model Predictive Control for tracking shows that the UVG achieves superior energy efficiency within its effective range while constituting a closed-form controller that requires fewer computational resources than numerical methods. The results highlight the UVG and related dynamics-based approaches as a compelling option for energy-conscious, task-focused robot designs. |
Nikolaos Leventakis, Panagiotis Vartholomeos Design, Control, and Experimental Verification of Exosuit for Assisting Knee Extension Conference IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), 2026. @conference{nokey,
title = {Design, Control, and Experimental Verification of Exosuit for Assisting Knee Extension},
author = {Nikolaos Leventakis, Panagiotis Vartholomeos},
url = {https://ieeexplore.ieee.org/document/11658009},
doi = {10.1109/AIM65483.2026.11658009},
year = {2026},
date = {2026-07-07},
urldate = {2026-07-07},
publisher = {IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM)},
abstract = {Incline walking and stair negotiation require more complex and more demanding effort than level walking. For this reason, the elderly population and people with mobility impairments are prone to falls and injuries when ascending or descending stairs and slopes. Although these activities are an integral part of daily living, only a few robotic assistive technologies address this need. The present paper is working towards this goal by presenting the concept, the computational model, the mechanical design, and the development of a soft robotic exoskeleton (exosuit) designed to assist knee extension during the ascend and descend of stairs. Experimental verification results are presented that demonstrate the functionality of the system for ascending tasks and show promising results in reducing muscle effort during stair negotiation.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Incline walking and stair negotiation require more complex and more demanding effort than level walking. For this reason, the elderly population and people with mobility impairments are prone to falls and injuries when ascending or descending stairs and slopes. Although these activities are an integral part of daily living, only a few robotic assistive technologies address this need. The present paper is working towards this goal by presenting the concept, the computational model, the mechanical design, and the development of a soft robotic exoskeleton (exosuit) designed to assist knee extension during the ascend and descend of stairs. Experimental verification results are presented that demonstrate the functionality of the system for ascending tasks and show promising results in reducing muscle effort during stair negotiation. |
Lydia Gialama, Panagiotis Kalozoumis, Ioannis Balatsoukas, Panagiotis Vartholomeos Force-Aware Computational Framework for Contact-Mediated Workspace Augmentation in Robotic Catheters Conference IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), 2026. @conference{nokey,
title = {Force-Aware Computational Framework for Contact-Mediated Workspace Augmentation in Robotic Catheters},
author = {Lydia Gialama, Panagiotis Kalozoumis, Ioannis Balatsoukas, Panagiotis Vartholomeos},
url = {https://ieeexplore.ieee.org/document/11657896},
doi = {10.1109/AIM65483.2026.11657896},
year = {2026},
date = {2026-07-07},
urldate = {2026-07-07},
publisher = {IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM)},
abstract = {Tendon-driven catheters are increasingly used in minimally invasive procedures involving the cardiovascular and neurovascular systems. However, their maneuverability remains extremely limited due to the narrow and tortuous anatomy through which they must navigate. Precise tip positioning in such confined spaces requires knowledge of the catheter’s reachable workspace and awareness of the contact forces exerted on the surrounding tissue, information that is difficult to obtain experimentally in real time. This work proposes a simulation-based framework that expands the reachable workspace of tendon-driven catheters in constrained environments and maps contact forces across it. A computationally efficient Lumped Parameter Model (LPM) is developed and deployed in MuJoCo to capture contact mechanics across 3,600 simulated scenarios covering a broad range of bending curvatures, obstacle offsets, and insertion lengths. The LPM is calibrated against a high-fidelity Finite Element Model and validated experimentally with tip position errors below 4% and force errors below 10% across all actuation cases. Jointly, the workspace and force maps establish the foundation for safe, force-aware catheter navigation among obstacles.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Tendon-driven catheters are increasingly used in minimally invasive procedures involving the cardiovascular and neurovascular systems. However, their maneuverability remains extremely limited due to the narrow and tortuous anatomy through which they must navigate. Precise tip positioning in such confined spaces requires knowledge of the catheter’s reachable workspace and awareness of the contact forces exerted on the surrounding tissue, information that is difficult to obtain experimentally in real time. This work proposes a simulation-based framework that expands the reachable workspace of tendon-driven catheters in constrained environments and maps contact forces across it. A computationally efficient Lumped Parameter Model (LPM) is developed and deployed in MuJoCo to capture contact mechanics across 3,600 simulated scenarios covering a broad range of bending curvatures, obstacle offsets, and insertion lengths. The LPM is calibrated against a high-fidelity Finite Element Model and validated experimentally with tip position errors below 4% and force errors below 10% across all actuation cases. Jointly, the workspace and force maps establish the foundation for safe, force-aware catheter navigation among obstacles. |
Alexandros Dimitriadis, George P. Moustris, Costas Tzafestas 2D Surgical Tool Pose Estimation via Top-Down Deep Learning and Adaptive Temporal Filtering Conference The 18th Hamlyn Symposium on Medical Robotics (HSMR'26), London, UK, 2026. @conference{nokey,
title = {2D Surgical Tool Pose Estimation via Top-Down Deep Learning and Adaptive Temporal Filtering},
author = {Alexandros Dimitriadis, George P. Moustris, Costas Tzafestas},
url = {https://www.researchgate.net/publication/404191853_2D_Surgical_Tool_Pose_Estimation_via_Top-Down_Deep_Learning_and_Adaptive_Temporal_Filtering},
doi = {10.31256/HSMR26.57},
year = {2026},
date = {2026-07-01},
urldate = {2026-07-01},
publisher = {The 18th Hamlyn Symposium on Medical Robotics (HSMR'26), London, UK},
abstract = {Accurate 2D surgical tool pose estimation is essential for intraoperative guidance and automated workflow analysis. In this work, we present a robust two-stage surgical pose estimation framework that combines a YOLOv8m detector with an RTMPose-m estimator. Our approach utilizes a coordinate classification head (SimCC) to achieve high accuracy across 14 instrument landmarks. To address the inherent instability of deep learning predictions in surgical environments, we introduce a confidence-weighted temporal filter. This adaptive mechanism dynamically adjusts smoothing intensity based on detection reliability, suppressing erroneous predictions and jitter while maintaining responsive tracking. Experimental results on the SurgPose dataset demonstrate high average precision across most tool geometries.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Accurate 2D surgical tool pose estimation is essential for intraoperative guidance and automated workflow analysis. In this work, we present a robust two-stage surgical pose estimation framework that combines a YOLOv8m detector with an RTMPose-m estimator. Our approach utilizes a coordinate classification head (SimCC) to achieve high accuracy across 14 instrument landmarks. To address the inherent instability of deep learning predictions in surgical environments, we introduce a confidence-weighted temporal filter. This adaptive mechanism dynamically adjusts smoothing intensity based on detection reliability, suppressing erroneous predictions and jitter while maintaining responsive tracking. Experimental results on the SurgPose dataset demonstrate high average precision across most tool geometries. |
Anargyros-Nektarios Skafidas, Aikaterini Smyrli, Evangelos Papadopoulos Exploiting Passive Dynamics for Steady-state Walking on Level and Inclined Terrain Conference Mechanism and Machine Theory Symposium (MMT Symposium), 2026. @conference{nokey,
title = {Exploiting Passive Dynamics for Steady-state Walking on Level and Inclined Terrain},
author = {Anargyros-Nektarios Skafidas, Aikaterini Smyrli, Evangelos Papadopoulos},
url = {https://www.researchgate.net/publication/408719476_Exploiting_passive_dynamics_for_steady-state_walking_on_level_and_inclined_terrain},
year = {2026},
date = {2026-06-21},
urldate = {2026-06-21},
publisher = {Mechanism and Machine Theory Symposium (MMT Symposium)},
abstract = {In passive dynamic walking, stable and energy-efficient bipedal gait is performed through the exploitation of the inherent dynamics of appropriately designed mechanical systems. However, purely passive bipeds are necessarily restricted to downhill walks, where gravity compensates for impacts and other losses. Extending such motions to level and uphill terrain while preserving the desirable characteristics of passive gait remains a significant challenge. Here, an under-actuated biped robot, designed to rely on and exploit its passive dynamics while in active locomotion, is studied for its ability to generate stable and energy-efficient locomotion on level and positively sloped ground.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
In passive dynamic walking, stable and energy-efficient bipedal gait is performed through the exploitation of the inherent dynamics of appropriately designed mechanical systems. However, purely passive bipeds are necessarily restricted to downhill walks, where gravity compensates for impacts and other losses. Extending such motions to level and uphill terrain while preserving the desirable characteristics of passive gait remains a significant challenge. Here, an under-actuated biped robot, designed to rely on and exploit its passive dynamics while in active locomotion, is studied for its ability to generate stable and energy-efficient locomotion on level and positively sloped ground. |
Aikaterini Smyrli, Evangelos Papadopoulos Morphological Intelligence in passive Dynamics: A Design Investigation on Biomechanically-inspired Bipedal Robots Conference Mechanism and Machine Theory Symposium (MMT Symposium), 2026. @conference{nokey,
title = {Morphological Intelligence in passive Dynamics: A Design Investigation on Biomechanically-inspired Bipedal Robots},
author = {Aikaterini Smyrli, Evangelos Papadopoulos},
url = {https://www.researchgate.net/publication/408719472_Morphological_Intelligence_in_Passive_Dynamics_A_design_investigation_of_biomechanically-inspired_bipedal_robots},
year = {2026},
date = {2026-06-21},
urldate = {2026-06-21},
publisher = {Mechanism and Machine Theory Symposium (MMT Symposium)},
abstract = {This work investigates the dynamics of bipedal locomotion through the modeling and progressive refinement of passive bio-inspired mechanical walkers, with emphasis on applications in robotics and lower-limb prosthetic systems. The study follows a structured mechanism-design progression, beginning from a minimal passive configuration and incrementally incorporating anatomically inspired elements that shape hybrid gait dynamics. The central hypothesis is that stable and energetically efficient walking behavior emerges primarily from mechanical morphology. By systematically augmenting structural features, quantitative relationships between geometry, mechanism design, inertial distribution, and passive gait characteristics are established.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
This work investigates the dynamics of bipedal locomotion through the modeling and progressive refinement of passive bio-inspired mechanical walkers, with emphasis on applications in robotics and lower-limb prosthetic systems. The study follows a structured mechanism-design progression, beginning from a minimal passive configuration and incrementally incorporating anatomically inspired elements that shape hybrid gait dynamics. The central hypothesis is that stable and energetically efficient walking behavior emerges primarily from mechanical morphology. By systematically augmenting structural features, quantitative relationships between geometry, mechanism design, inertial distribution, and passive gait characteristics are established. |
Konstantinos Chaldaiopoulos, Niki Efthymiou, Athanasia Zlatintsi, Panagiotis Filntisis, Katerina Pastra, Petros Maragos PredJSal: Video Saliency via Predictive Self-Supervised Representations Conference IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), NTIRE Workshop, 2026. @conference{nokey,
title = {PredJSal: Video Saliency via Predictive Self-Supervised Representations},
author = {Konstantinos Chaldaiopoulos, Niki Efthymiou, Athanasia Zlatintsi, Panagiotis Filntisis, Katerina Pastra, Petros Maragos},
url = {https://openaccess.thecvf.com/content/CVPR2026W/NTIRE/papers/Chaldaiopoulos_PredJSal_Video_Saliency_via_Predictive_Self-Supervised_Representations_CVPRW_2026_paper.pdf},
year = {2026},
date = {2026-06-03},
urldate = {2026-06-03},
publisher = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), NTIRE Workshop},
abstract = {Predicting where humans look in video requires models that go beyond appearance statistics and capture the dynamic, physics-driven nature of visual attention. We argue that self-supervised video representations trained under a world-predictive objective are a theoretically principled substrate for this task, as they align with the prediction error mechanisms that underlie biological saliency under predictive coding theory. To this end, we propose PredJSal, a video saliency framework built on V-JEPA2 augmented with temporal self-attention and a multi-scale 3D convolutional decoder. Relying solely on self-supervised joint embedding predictive pretraining, PredJSal ranks second at the NTIRE 2026 Video Saliency Prediction Challenge with an identical mean rank to the first-placed entry, and sets a new state of the art on the DHF1K dataset. Our code and weights will be available at https://github.com/KChald/PredJSal.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Predicting where humans look in video requires models that go beyond appearance statistics and capture the dynamic, physics-driven nature of visual attention. We argue that self-supervised video representations trained under a world-predictive objective are a theoretically principled substrate for this task, as they align with the prediction error mechanisms that underlie biological saliency under predictive coding theory. To this end, we propose PredJSal, a video saliency framework built on V-JEPA2 augmented with temporal self-attention and a multi-scale 3D convolutional decoder. Relying solely on self-supervised joint embedding predictive pretraining, PredJSal ranks second at the NTIRE 2026 Video Saliency Prediction Challenge with an identical mean rank to the first-placed entry, and sets a new state of the art on the DHF1K dataset. Our code and weights will be available at https://github.com/KChald/PredJSal. |
Panagiotis P. Filntisis, George Retsinas, Radek Daněček, Vanessa Sklyarova, Petros Maragos, Timo Bolkart Registration-Free Learnable Multi-View Capture of Faces in Dense Semantic Correspondence Conference IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. @conference{nokey,
title = {Registration-Free Learnable Multi-View Capture of Faces in Dense Semantic Correspondence},
author = {Panagiotis P. Filntisis, George Retsinas, Radek Daněček, Vanessa Sklyarova, Petros Maragos, Timo Bolkart},
url = {https://openaccess.thecvf.com/content/CVPR2026/papers/Filntisis_Registration-Free_Learnable_Multi-View_Capture_of_Faces_in_Dense_Semantic_Correspondence_CVPR_2026_paper.pdf},
doi = {10.48550/arXiv.2605.01450},
year = {2026},
date = {2026-06-02},
urldate = {2026-06-02},
pages = {14512-14523},
publisher = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
abstract = {Recent frameworks like ToFu and TEMPEH provide an automated alternative to classical registration pipelines by predicting 3D meshes in dense semantic correspondence directly from calibrated multi-view images. However, these learning-based methods rely on the slow, manual registration pipelines they aim to replace for their training supervision. We overcome this limitation with MOCHI (Multiview Optimizable Correspondence of Heads from Images), a multi-view 3D face prediction framework trained without requiring registered training data. MOCHI eliminates the registration data dependency by enforcing topological consistency through a pseudo-linear inverse kinematic solver. Semantic alignment is guided by dense keypoints from a 2D landmark predictor trained exclusively on synthetic data. Our analysis further reveals that standard point-to-surface distances induce training instabilities and visual artifacts in registration-free settings. We propose pointmap- and normal-based losses instead, which provide smoother gradients and superior reconstruction fidelity. Finally, we introduce a test-time optimization scheme that refines network weights over a few dozen iterations. This approach bridges the gap between feed-forward efficiency and iterative optimization precision, allowing MOCHI to outperform traditional labor-intensive pipelines in both reconstruction accuracy and visual quality. Code and model are public at: https://filby89.github.io/mochi.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Recent frameworks like ToFu and TEMPEH provide an automated alternative to classical registration pipelines by predicting 3D meshes in dense semantic correspondence directly from calibrated multi-view images. However, these learning-based methods rely on the slow, manual registration pipelines they aim to replace for their training supervision. We overcome this limitation with MOCHI (Multiview Optimizable Correspondence of Heads from Images), a multi-view 3D face prediction framework trained without requiring registered training data. MOCHI eliminates the registration data dependency by enforcing topological consistency through a pseudo-linear inverse kinematic solver. Semantic alignment is guided by dense keypoints from a 2D landmark predictor trained exclusively on synthetic data. Our analysis further reveals that standard point-to-surface distances induce training instabilities and visual artifacts in registration-free settings. We propose pointmap- and normal-based losses instead, which provide smoother gradients and superior reconstruction fidelity. Finally, we introduce a test-time optimization scheme that refines network weights over a few dozen iterations. This approach bridges the gap between feed-forward efficiency and iterative optimization precision, allowing MOCHI to outperform traditional labor-intensive pipelines in both reconstruction accuracy and visual quality. Code and model are public at: https://filby89.github.io/mochi. |
Nikolas Chatzis, Ioannis Kordonis, Emmanouil Theodosis, Petros Maragos Shared-Weights Extender and Gradient Voting for Neural Network Expansion Conference IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026. @conference{nokey,
title = {Shared-Weights Extender and Gradient Voting for Neural Network Expansion},
author = {Nikolas Chatzis, Ioannis Kordonis, Emmanouil Theodosis, Petros Maragos},
url = {https://ieeexplore.ieee.org/document/11462494},
doi = {10.1109/ICASSP55912.2026.11462494},
year = {2026},
date = {2026-05-03},
urldate = {2026-05-03},
publisher = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
abstract = {Expanding neural networks during training is a promising way to augment capacity without retraining larger models from scratch. However, newly added neurons often fail to adjust to a trained network and become inactive, providing no contribution to capacity growth. We propose the Shared-Weights Extender (SWE), a novel method explicitly designed to prevent inactivity of new neurons by coupling them with existing ones for smooth integration. In parallel, we introduce the Steepest Voting Distributor (SVoD), a gradient-based method for allocating neurons across layers during deep network expansion. Our extensive benchmarking on four datasets shows that our method can effectively suppress neuron inactivity and achieve better performance compared to other expanding methods and baselines.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Expanding neural networks during training is a promising way to augment capacity without retraining larger models from scratch. However, newly added neurons often fail to adjust to a trained network and become inactive, providing no contribution to capacity growth. We propose the Shared-Weights Extender (SWE), a novel method explicitly designed to prevent inactivity of new neurons by coupling them with existing ones for smooth integration. In parallel, we introduce the Steepest Voting Distributor (SVoD), a gradient-based method for allocating neurons across layers during deep network expansion. Our extensive benchmarking on four datasets shows that our method can effectively suppress neuron inactivity and achieve better performance compared to other expanding methods and baselines. |
Dimitris N. Makropoulos, Christos Garoufis, Antigoni Tsiami, Panagiotis P. Filntisis, Petros Maragos Representation-Diverse Self-Supervision for Cross-Domain Bioacoustic Learning in Low-Resource Settings Conference IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026. @conference{nokey,
title = {Representation-Diverse Self-Supervision for Cross-Domain Bioacoustic Learning in Low-Resource Settings},
author = {Dimitris N. Makropoulos, Christos Garoufis, Antigoni Tsiami, Panagiotis P. Filntisis, Petros Maragos},
url = {https://ieeexplore.ieee.org/document/11461507},
doi = {10.1109/ICASSP55912.2026.11461507},
year = {2026},
date = {2026-05-03},
urldate = {2026-05-03},
publisher = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
abstract = {We propose a contrastive self-supervised learning framework to improve transfer learning across dissimilar bioacoustic datasets. Amplitude - and frequency - modulated (AM-FM) delphinid vocalizations from the Watkins Marine Mammal Sound Database (WMMSD) are used for pretraining, while two bird-call datasets -Rain Forest Connection Species Audio Detection (RFCx - bird and frog calls) and BirdCLEF 2022 (bird vocalizations)-serve for downstream evaluation. We train ResNet18, MobileNetV2, and Vision Transformer (ViT-B/16) backbones, in a two-stage pipeline. In the pretraining stage, the model contrasts different time-frequency representations of the same audio, such as spectrogram-energogram pairs derived via the discrete or regularized Teager-Kaiser Energy Operator (TKEO). In the fine-tuning stage, the pretrained models are adapted for bird vocalization classification. Experiments show that representation-diverse contrastive learning on delphinid calls, followed by fine-tuning, consistently outperforms both fully supervised baselines and conventional self-supervised approaches based solely on augmented spectrograms. The best results are obtained when contrasting spectrograms with TKEO-based energograms, indicating that cross-representation contrastive learning yields more robust and transferable embeddings for bioacoustic tasks.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
We propose a contrastive self-supervised learning framework to improve transfer learning across dissimilar bioacoustic datasets. Amplitude - and frequency - modulated (AM-FM) delphinid vocalizations from the Watkins Marine Mammal Sound Database (WMMSD) are used for pretraining, while two bird-call datasets -Rain Forest Connection Species Audio Detection (RFCx - bird and frog calls) and BirdCLEF 2022 (bird vocalizations)-serve for downstream evaluation. We train ResNet18, MobileNetV2, and Vision Transformer (ViT-B/16) backbones, in a two-stage pipeline. In the pretraining stage, the model contrasts different time-frequency representations of the same audio, such as spectrogram-energogram pairs derived via the discrete or regularized Teager-Kaiser Energy Operator (TKEO). In the fine-tuning stage, the pretrained models are adapted for bird vocalization classification. Experiments show that representation-diverse contrastive learning on delphinid calls, followed by fine-tuning, consistently outperforms both fully supervised baselines and conventional self-supervised approaches based solely on augmented spectrograms. The best results are obtained when contrasting spectrograms with TKEO-based energograms, indicating that cross-representation contrastive learning yields more robust and transferable embeddings for bioacoustic tasks. |
Georgios Chatzichristodoulou, Niki Efthymiou, Panagiotis Filntisis, Georgios Pavlakos, Petros Maragos Age-Inclusive 3D Human Mesh Recovery for Action-Preserving Data Anonymization Conference IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026. @conference{nokey,
title = {Age-Inclusive 3D Human Mesh Recovery for Action-Preserving Data Anonymization},
author = {Georgios Chatzichristodoulou, Niki Efthymiou, Panagiotis Filntisis, Georgios Pavlakos, Petros Maragos},
url = {https://openaccess.thecvf.com/content/CVPR2026W/CV4CHL/papers/Chatzichristodoulou_Age-Inclusive_3D_Human_Mesh_Recovery_for_Action-Preserving_Data_Anonymization_CVPRW_2026_paper.pdf},
doi = {10.48550/arXiv.2512.05259},
year = {2026},
date = {2026-01-04},
urldate = {2026-01-04},
pages = {5553-5563},
publisher = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
abstract = {While three-dimensional (3D) human shape and pose estimation is a highly researched area that has yielded significant advances, the resulting methods, despite performing well for the adult population, generally fail to generalize effectively to children and infants. This paper addresses this challenge by introducing AionHMR, a comprehensive framework designed to bridge this domain gap. We propose an optimization-based method that extends a top-performing model by incorporating the SMPL-A body model, enabling the concurrent and accurate modeling of adults, children, and infants. Leveraging this approach, we generated pseudo-ground-truth annotations for publicly available child and infant image databases. Using these new training data, we then developed and trained a specialized transformer-based deep learning model capable of real-time 3D age-inclusive human reconstruction. Extensive experiments demonstrate that our methods significantly improve shape and pose estimation for children and infants without compromising accuracy on adults. Importantly, our reconstructed meshes serve as privacy-preserving substitutes for raw images, retaining essential action, pose, and geometry information while enabling anonymized datasets release. As a demonstration, we introduce the 3D-BabyRobot dataset, a collection of action-preserving 3D reconstructions of children interacting with robots. This work bridges a crucial domain gap and establishes a foundation for inclusive, privacy-aware, and age-diverse 3D human modeling.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
While three-dimensional (3D) human shape and pose estimation is a highly researched area that has yielded significant advances, the resulting methods, despite performing well for the adult population, generally fail to generalize effectively to children and infants. This paper addresses this challenge by introducing AionHMR, a comprehensive framework designed to bridge this domain gap. We propose an optimization-based method that extends a top-performing model by incorporating the SMPL-A body model, enabling the concurrent and accurate modeling of adults, children, and infants. Leveraging this approach, we generated pseudo-ground-truth annotations for publicly available child and infant image databases. Using these new training data, we then developed and trained a specialized transformer-based deep learning model capable of real-time 3D age-inclusive human reconstruction. Extensive experiments demonstrate that our methods significantly improve shape and pose estimation for children and infants without compromising accuracy on adults. Importantly, our reconstructed meshes serve as privacy-preserving substitutes for raw images, retaining essential action, pose, and geometry information while enabling anonymized datasets release. As a demonstration, we introduce the 3D-BabyRobot dataset, a collection of action-preserving 3D reconstructions of children interacting with robots. This work bridges a crucial domain gap and establishes a foundation for inclusive, privacy-aware, and age-diverse 3D human modeling. |
2025
|
Siyuan Li, Shikai Fang, Lei Cheng, Feng Yin, Yik-Chung Wu, Peter Gerstoft, Sergios Theodoridis When Bayesian Tensor Completion Meets Multioutput Gaussian Processes: Functional Universality and Rank Learning Journal Article In: IEEE Transactions on Signal Processing , vol. 73, pp. 5319 - 5335, 2025. @article{nokey,
title = {When Bayesian Tensor Completion Meets Multioutput Gaussian Processes: Functional Universality and Rank Learning},
author = {Siyuan Li, Shikai Fang, Lei Cheng, Feng Yin, Yik-Chung Wu, Peter Gerstoft, Sergios Theodoridis},
url = {https://ieeexplore.ieee.org/document/11271811},
doi = {10.1109/TSP.2025.3639229},
year = {2025},
date = {2025-12-02},
urldate = {2025-12-02},
journal = {IEEE Transactions on Signal Processing },
volume = {73},
pages = {5319 - 5335},
abstract = {Functional tensor decomposition can analyze multi-dimensional data with real-valued indices, paving the path for applications in machine learning and signal processing. A limitation of existing approaches is the assumption that the tensor rank—a critical parameter governing model complexity—is known. However, determining the optimal rank is a non-deterministic polynomial-time hard (NP-hard) task and there is a limited understanding regarding the expressive power of functional low-rank tensor models for continuous signals. We propose a rank-revealing functional Bayesian tensor completion (RR-FBTC) method. Modeling the latent functions through carefully designed multioutput Gaussian processes, RR-FBTC handles tensors with real-valued indices while enabling automatic tensor rank determination during the inference process. We establish the universal approximation property of the model for continuous multi-dimensional signals, demonstrating its expressive power in a concise format. To learn this model, we employ the variational inference framework and derive an efficient algorithm with closed-form updates. Experiments on both synthetic and real-world datasets demonstrate the effectiveness and superiority of the RR-FBTC over state-of-the-art approaches. The code is available at https://github.com/OceanSTARLab/RR-FBTC.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Functional tensor decomposition can analyze multi-dimensional data with real-valued indices, paving the path for applications in machine learning and signal processing. A limitation of existing approaches is the assumption that the tensor rank—a critical parameter governing model complexity—is known. However, determining the optimal rank is a non-deterministic polynomial-time hard (NP-hard) task and there is a limited understanding regarding the expressive power of functional low-rank tensor models for continuous signals. We propose a rank-revealing functional Bayesian tensor completion (RR-FBTC) method. Modeling the latent functions through carefully designed multioutput Gaussian processes, RR-FBTC handles tensors with real-valued indices while enabling automatic tensor rank determination during the inference process. We establish the universal approximation property of the model for continuous multi-dimensional signals, demonstrating its expressive power in a concise format. To learn this model, we employ the variational inference framework and derive an efficient algorithm with closed-form updates. Experiments on both synthetic and real-world datasets demonstrate the effectiveness and superiority of the RR-FBTC over state-of-the-art approaches. The code is available at https://github.com/OceanSTARLab/RR-FBTC. |
Bill Psomas, George Retsinas, Nikos Efthymiadis, Panagiotis Filntisis, Yannis Avrithis, Petros Maragos, Ondrej Chum, Giorgos Tolias Instance-Level Composed Image Retrieval Conference Conference on Neural Information Processing Systems (NeurIPS), 2025. @conference{nokey,
title = {Instance-Level Composed Image Retrieval},
author = {Bill Psomas, George Retsinas, Nikos Efthymiadis, Panagiotis Filntisis, Yannis Avrithis, Petros Maragos, Ondrej Chum, Giorgos Tolias},
url = {https://vrg.fel.cvut.cz/icir/},
doi = {10.52202/085713-2315},
year = {2025},
date = {2025-12-02},
urldate = {2025-11-30},
publisher = {Conference on Neural Information Processing Systems (NeurIPS)},
abstract = {The progress of composed image retrieval (CIR), a popular research direction in image retrieval, where a combined visual and textual query is used, is held back by the absence of high-quality training and evaluation data. We introduce a new evaluation dataset, i-CIR, which, unlike existing datasets, focuses on an instance-level class definition. The goal is to retrieve images that contain the same particular object as the visual query, presented under a variety of modifications defined by textual queries. Its design and curation process keep the dataset compact to facilitate future research, while maintaining its challenge—comparable to retrieval among more than 40M random distractors—through a semi-automated selection of hard negatives. To overcome the challenge of obtaining clean, diverse, and suitable training data, we leverage pre-trained vision-and-language models (VLMs) in a training-free approach called BASIC. The method separately estimates query-image-to-image and query-text-to-image similarities, performing late fusion to upweight images that satisfy both queries, while down-weighting those that exhibit high similarity with only one of the two. Each individual similarity is further improved by a set of components that are simple and intuitive. BASIC sets a new state of the art on i-CIR but also on existing CIR datasets that follow a semantic-level class definition. Project page: https://vrg.fel.cvut.cz/icir/.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
The progress of composed image retrieval (CIR), a popular research direction in image retrieval, where a combined visual and textual query is used, is held back by the absence of high-quality training and evaluation data. We introduce a new evaluation dataset, i-CIR, which, unlike existing datasets, focuses on an instance-level class definition. The goal is to retrieve images that contain the same particular object as the visual query, presented under a variety of modifications defined by textual queries. Its design and curation process keep the dataset compact to facilitate future research, while maintaining its challenge—comparable to retrieval among more than 40M random distractors—through a semi-automated selection of hard negatives. To overcome the challenge of obtaining clean, diverse, and suitable training data, we leverage pre-trained vision-and-language models (VLMs) in a training-free approach called BASIC. The method separately estimates query-image-to-image and query-text-to-image similarities, performing late fusion to upweight images that satisfy both queries, while down-weighting those that exhibit high similarity with only one of the two. Each individual similarity is further improved by a set of components that are simple and intuitive. BASIC sets a new state of the art on i-CIR but also on existing CIR datasets that follow a semantic-level class definition. Project page: https://vrg.fel.cvut.cz/icir/. |
Richard Cornelius Suwandi, Feng Yin, Juntao Wang, Renjie Li, Tsung-Hui Chang, Sergios Theodoridis Adaptive Kernel Design for Bayesian Optimization is a Piece of CAKE with LLMs Conference Conference on Neural Information Processing Systems (NeurIPS), 2025. @conference{nokey,
title = {Adaptive Kernel Design for Bayesian Optimization is a Piece of CAKE with LLMs},
author = {Richard Cornelius Suwandi, Feng Yin, Juntao Wang, Renjie Li, Tsung-Hui Chang, Sergios Theodoridis},
url = {https://neurips.cc/virtual/2025/loc/san-diego/poster/119378},
doi = {10.48550/arXiv.2509.17998},
year = {2025},
date = {2025-12-02},
urldate = {2025-11-30},
publisher = {Conference on Neural Information Processing Systems (NeurIPS)},
abstract = {The efficiency of Bayesian optimization (BO) relies heavily on the choice of the Gaussian process (GP) kernel, which plays a central role in balancing exploration and exploitation under limited evaluation budgets. Traditional BO methods often rely on fixed or heuristic kernel selection strategies, which can result in slow convergence or suboptimal solutions when the chosen kernel is poorly suited to the underlying objective function. To address this limitation, we propose a freshly-baked Context-Aware Kernel Evolution (CAKE) to enhance BO with large language models (LLMs). Concretely, CAKE leverages LLMs as the crossover and mutation operators to adaptively generate and refine GP kernels based on the observed data throughout the optimization process. To maximize the power of CAKE, we further propose BIC-Acquisition Kernel Ranking (BAKER) to select the most effective kernel through balancing the model fit measured by the Bayesian information criterion (BIC) with the expected improvement at each iteration of BO. Extensive experiments demonstrate that our fresh CAKE-based BO method consistently outperforms established baselines across a range of real-world tasks, including hyperparameter optimization, controller tuning, and photonic chip design. Our code is publicly available at https://github.com/richardcsuwandi/cake.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
The efficiency of Bayesian optimization (BO) relies heavily on the choice of the Gaussian process (GP) kernel, which plays a central role in balancing exploration and exploitation under limited evaluation budgets. Traditional BO methods often rely on fixed or heuristic kernel selection strategies, which can result in slow convergence or suboptimal solutions when the chosen kernel is poorly suited to the underlying objective function. To address this limitation, we propose a freshly-baked Context-Aware Kernel Evolution (CAKE) to enhance BO with large language models (LLMs). Concretely, CAKE leverages LLMs as the crossover and mutation operators to adaptively generate and refine GP kernels based on the observed data throughout the optimization process. To maximize the power of CAKE, we further propose BIC-Acquisition Kernel Ranking (BAKER) to select the most effective kernel through balancing the model fit measured by the Bayesian information criterion (BIC) with the expected improvement at each iteration of BO. Extensive experiments demonstrate that our fresh CAKE-based BO method consistently outperforms established baselines across a range of real-world tasks, including hyperparameter optimization, controller tuning, and photonic chip design. Our code is publicly available at https://github.com/richardcsuwandi/cake. |
Katerina Papadimitriou, Panagiotis Filntisis, George Retsinas, Gerasimos Potamianos, Petros Maragos Seeing in 2D, Thinking in 3D: 3D Hand Mesh-Guided Feature Learning for Continuous Fingerspelling Conference EEE/CVF International Conference on Computer Vision, ICCV (ICCVW), 2025. @conference{nokey,
title = {Seeing in 2D, Thinking in 3D: 3D Hand Mesh-Guided Feature Learning for Continuous Fingerspelling},
author = {Katerina Papadimitriou, Panagiotis Filntisis, George Retsinas, Gerasimos Potamianos, Petros Maragos},
url = {https://openaccess.thecvf.com/content/ICCV2025W/CV4A11y/papers/Papadimitriou_Seeing_in_2D_Thinking_in_3D_3D_Hand_Mesh-Guided_Feature_ICCVW_2025_paper.pdf},
doi = {10.1109/ICCVW69036.2025.00695},
year = {2025},
date = {2025-10-19},
urldate = {2025-10-19},
publisher = {EEE/CVF International Conference on Computer Vision, ICCV (ICCVW)},
abstract = {Recognizing continuous fingerspelling from monocular RGB video is a highly challenging task due to complex hand articulation, coarticulation effects, and significant intersigner variability. Prior methods use either raw visual features, which lack structural awareness of fine-grained finger dynamics, or parallel RGBpose streams from explicit pose estimation, which add substantial inference-time overhead. In this work, we propose a novel knowledge distillation framework that transfers rich hand articulation knowledge from HAMER, a foundation model for 3D hand mesh/pose reconstruction, into a lightweight, RGB-only fingerspelling recognizer. We extract high-level pose embeddings from HAMER's Transformer head, which encode detailed hand structure, and distill them into a ResNet34-based appearance encoder via a dedicated training objective. Subsequently, the learned pose-aware features are fed into a ID-CNN and BiGRU for temporal modeling, with the full system trained using both connectionist temporal classification (CTC) and a knowledge distillation loss. Notably, our approach does not rely on the teacher model (HAMER) at inference time, thus enabling real-time performance. We evaluate our method on two American sign language (ASL) benchmark-fingerspelling datasets, as well as a studio-quality Greek fingerspelling corpus. Our model achieves state-of-the-art accuracy with over 3x lower inference time than prior methods, offering an effective trade-off between accuracy and efficiency for real-time deployment.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Recognizing continuous fingerspelling from monocular RGB video is a highly challenging task due to complex hand articulation, coarticulation effects, and significant intersigner variability. Prior methods use either raw visual features, which lack structural awareness of fine-grained finger dynamics, or parallel RGBpose streams from explicit pose estimation, which add substantial inference-time overhead. In this work, we propose a novel knowledge distillation framework that transfers rich hand articulation knowledge from HAMER, a foundation model for 3D hand mesh/pose reconstruction, into a lightweight, RGB-only fingerspelling recognizer. We extract high-level pose embeddings from HAMER's Transformer head, which encode detailed hand structure, and distill them into a ResNet34-based appearance encoder via a dedicated training objective. Subsequently, the learned pose-aware features are fed into a ID-CNN and BiGRU for temporal modeling, with the full system trained using both connectionist temporal classification (CTC) and a knowledge distillation loss. Notably, our approach does not rely on the teacher model (HAMER) at inference time, thus enabling real-time performance. We evaluate our method on two American sign language (ASL) benchmark-fingerspelling datasets, as well as a studio-quality Greek fingerspelling corpus. Our model achieves state-of-the-art accuracy with over 3x lower inference time than prior methods, offering an effective trade-off between accuracy and efficiency for real-time deployment. |
Tilemachos Aravanis, Panagiotis Filntisis, Petros Maragos, George Retsinas Only-Style: Stylistic Consistency in Image Generation without Content Leakage Conference IEEE/CVF International Conference on Computer Vision, ICCV 2025 (ICCVW), 2025. @conference{nokey,
title = {Only-Style: Stylistic Consistency in Image Generation without Content Leakage},
author = {Tilemachos Aravanis, Panagiotis Filntisis, Petros Maragos, George Retsinas },
url = {https://openaccess.thecvf.com/content/ICCV2025W/P13N/papers/Aravanis_Only-Style_Stylistic_Consistency_in_Image_Generation_without_Content_Leakage_ICCVW_2025_paper.pdf},
doi = {10.1109/ICCVW69036.2025.00723},
year = {2025},
date = {2025-10-19},
urldate = {2025-10-19},
publisher = {IEEE/CVF International Conference on Computer Vision, ICCV 2025 (ICCVW)},
abstract = {Generating images in a consistent reference visual style remains a challenging computer vision task. State-of-the art methods aiming for style-consistent generation struggle to effectively separate semantic content from stylistic elements, leading to content leakage from the image provided as a reference to the targets. To address this challenge, we propose Only-Style: a method designed to mitigate content leakage in a semantically coherent manner while preserving stylistic consistency. Only-Style works by localizing content leakage during inference, allowing the adaptive tuning of a parameter that controls the style alignment process, specifically within the image patches containing the subject in the reference image. This adaptive process best balances stylistic consistency with leakage elimination. Moreover, the localization of content leakage can function as a standalone component, given a reference-target image pair, allowing the adaptive tuning of any method-specific parameter that provides control over the impact of the stylistic reference. In addition, we propose a novel evaluation framework to quantify the success of style-consistent generations in avoiding undesired content leakage. Our approach demonstrates a significant improvement over state-of-theart methods through extensive evaluation across diverse instances, consistently achieving robust stylistic consistency without undesired content leakage. },
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Generating images in a consistent reference visual style remains a challenging computer vision task. State-of-the art methods aiming for style-consistent generation struggle to effectively separate semantic content from stylistic elements, leading to content leakage from the image provided as a reference to the targets. To address this challenge, we propose Only-Style: a method designed to mitigate content leakage in a semantically coherent manner while preserving stylistic consistency. Only-Style works by localizing content leakage during inference, allowing the adaptive tuning of a parameter that controls the style alignment process, specifically within the image patches containing the subject in the reference image. This adaptive process best balances stylistic consistency with leakage elimination. Moreover, the localization of content leakage can function as a standalone component, given a reference-target image pair, allowing the adaptive tuning of any method-specific parameter that provides control over the impact of the stylistic reference. In addition, we propose a novel evaluation framework to quantify the success of style-consistent generations in avoiding undesired content leakage. Our approach demonstrates a significant improvement over state-of-theart methods through extensive evaluation across diverse instances, consistently achieving robust stylistic consistency without undesired content leakage. |
Joana Plewnia, Tamim Asfour Combining Episodic Memory and LLMs for the Verbalization of Robot Experiences Conference IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids), 2025. @conference{nokey,
title = {Combining Episodic Memory and LLMs for the Verbalization of Robot Experiences},
author = {Joana Plewnia, Tamim Asfour},
url = {https://h2t.iar.kit.edu/pdf/Plewnia2025b.pdf},
doi = {10.1109/HUMANOIDS65713.2025.11203157},
year = {2025},
date = {2025-09-30},
urldate = {2025-09-30},
publisher = {IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids)},
abstract = {The ability to communicate past experiences is fundamental for intelligent and natural interaction. Humanoid robots continuously accumulate rich, multi-modal experiential data and must be able to articulate their episodic experiences in natural language to support effective human-robot communication. Existing approaches either rely on the generalization capabilities of large language models (LLMs), sometimes combined with episodic memory, or the precision of rule-based verbalization systems, each presenting limitations when used in isolation. In this work, we present a novel hybrid framework that integrates the adaptability of LLMs with the robustness of rule-based methods and the generalizable structure of memory-based approaches. Our system implements strategies to retrieve and transform memory representations of past perceptions and actions into natural language responses. This enables humanoid robots to respond to natural language queries about their experience. Experimental evaluation based on a set of distinct query types demonstrates that our approach successfully answers 89.4 % of episodic memory questions with human-in-the-loop refinement, while reducing token consumption by 97 % compared to pure LLM-based methods. Furthermore, we demonstrate the system's extensibility by leveraging LLMs, such as GPT-4.1, to expand the range of permissible queries through example-based interaction. The evaluation on our humanoid robot Armar-7 performing household tasks validates that our hybrid approach balances response quality with computational efficiency to address the crucial need for dependable yet flexible verbalization of robot experiences. Code and examples are available at robotexperienceverbalization.github.io/REV.github.io/},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
The ability to communicate past experiences is fundamental for intelligent and natural interaction. Humanoid robots continuously accumulate rich, multi-modal experiential data and must be able to articulate their episodic experiences in natural language to support effective human-robot communication. Existing approaches either rely on the generalization capabilities of large language models (LLMs), sometimes combined with episodic memory, or the precision of rule-based verbalization systems, each presenting limitations when used in isolation. In this work, we present a novel hybrid framework that integrates the adaptability of LLMs with the robustness of rule-based methods and the generalizable structure of memory-based approaches. Our system implements strategies to retrieve and transform memory representations of past perceptions and actions into natural language responses. This enables humanoid robots to respond to natural language queries about their experience. Experimental evaluation based on a set of distinct query types demonstrates that our approach successfully answers 89.4 % of episodic memory questions with human-in-the-loop refinement, while reducing token consumption by 97 % compared to pure LLM-based methods. Furthermore, we demonstrate the system's extensibility by leveraging LLMs, such as GPT-4.1, to expand the range of permissible queries through example-based interaction. The evaluation on our humanoid robot Armar-7 performing household tasks validates that our hybrid approach balances response quality with computational efficiency to address the crucial need for dependable yet flexible verbalization of robot experiences. Code and examples are available at robotexperienceverbalization.github.io/REV.github.io/ |
Tim Wolk, Charlotte Marquardt, Miha Dežman, Tamim Asfour Comparing Machine Learning Methods for Force Myography Based Estimation of Isokinetic Knee and Ankle Torques Conference IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids), 2025. @conference{nokey,
title = {Comparing Machine Learning Methods for Force Myography Based Estimation of Isokinetic Knee and Ankle Torques},
author = {Tim Wolk, Charlotte Marquardt, Miha Dežman, Tamim Asfour},
url = {https://ieeexplore.ieee.org/abstract/document/11203135},
doi = {10.1109/HUMANOIDS65713.2025.11203135},
year = {2025},
date = {2025-09-30},
urldate = {2025-09-30},
publisher = {IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids)},
abstract = {Wearable sensors enable accurate estimation of joint moments through easy-to-use myography-based methods, such as force myography (FMG), offering practical benefits and valuable insights into continuous muscle state estimation to enhance control strategies. This paper presents a comparative analysis of four commonly used machine learning methods, Gaussian process regression (GPR), support vector regression (SVR), feed-forward neural network (FFNN), and temporal convolutional network (TCN), for estimation of human knee and ankle joint torques based on joint angles, velocities, and FMG signals from eight muscles on the human leg. The performance of the methods was evaluated on isokinetic motions of ten participants and compared to the models enhanced by electromyography (EMG) signals. Among the evaluated models, neural networks consistently demonstrated the highest accuracy in both inter- and intra-participant validations. Incorporating FMG modality yielded comparable performance to EMGbased estimation for unknown participants. Additionally, FMG outperforms EMG-based estimation in novel task characteristics within a single participant. These findings demonstrate the potential of FMG as a viable alternative to EMG for human joint torque estimation and highlight its potential for personalized exoskeleton control.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Wearable sensors enable accurate estimation of joint moments through easy-to-use myography-based methods, such as force myography (FMG), offering practical benefits and valuable insights into continuous muscle state estimation to enhance control strategies. This paper presents a comparative analysis of four commonly used machine learning methods, Gaussian process regression (GPR), support vector regression (SVR), feed-forward neural network (FFNN), and temporal convolutional network (TCN), for estimation of human knee and ankle joint torques based on joint angles, velocities, and FMG signals from eight muscles on the human leg. The performance of the methods was evaluated on isokinetic motions of ten participants and compared to the models enhanced by electromyography (EMG) signals. Among the evaluated models, neural networks consistently demonstrated the highest accuracy in both inter- and intra-participant validations. Incorporating FMG modality yielded comparable performance to EMGbased estimation for unknown participants. Additionally, FMG outperforms EMG-based estimation in novel task characteristics within a single participant. These findings demonstrate the potential of FMG as a viable alternative to EMG for human joint torque estimation and highlight its potential for personalized exoskeleton control. |
Orestis Konstantaropoulos, Mehdi Khamassi, Petros Maragos, George Retsinas Push, See, Predict: Emergent Perception Through Intrinsically Motivated Play Conference IEEE International Conference on Development and Learning (ICDL), 2025. @conference{nokey,
title = {Push, See, Predict: Emergent Perception Through Intrinsically Motivated Play},
author = {Orestis Konstantaropoulos, Mehdi Khamassi, Petros Maragos, George Retsinas},
url = {https://ieeexplore.ieee.org/document/11204356},
doi = {10.1109/ICDL63968.2025.11204356},
year = {2025},
date = {2025-09-16},
urldate = {2025-09-16},
publisher = {IEEE International Conference on Development and Learning (ICDL)},
abstract = {Unlike conventional vision systems that rely on passive observation, biological agents learn through physical interaction. Can a robot similarly develop an understanding of its environment purely through interaction, without prior knowledge or external supervision? In this work, we explore how artificial agents can autonomously learn via intrinsic motivation, much like how children engage in curious free play. We propose a novel, fully self-supervised, object-centric learning framework. The system first segments visual input into discrete entities using Slot Attention, trained on data collected from random robotic actions. A graph-based world model is then trained to predict object-centric dynamics but initially struggles to capture object motion due to the limited diversity of the initial interactions. To overcome this, we introduce an intrinsically motivated reward signal based on world model’s prediction error, which drives a policy to collect more informative trajectories. This results in up to three times more object displacement than random actions, significantly enriching the dataset. Fine-tuning both the vision and world model on these data improves prediction and reconstruction performance. We validate our method in a simulated robotic environment with diverse objects, demonstrating that meaningful visual and physical representations can emerge entirely from self-supervised interaction. This highlights the potential of intrinsically motivated, object-centric learning for autonomous world perception and modeling.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Unlike conventional vision systems that rely on passive observation, biological agents learn through physical interaction. Can a robot similarly develop an understanding of its environment purely through interaction, without prior knowledge or external supervision? In this work, we explore how artificial agents can autonomously learn via intrinsic motivation, much like how children engage in curious free play. We propose a novel, fully self-supervised, object-centric learning framework. The system first segments visual input into discrete entities using Slot Attention, trained on data collected from random robotic actions. A graph-based world model is then trained to predict object-centric dynamics but initially struggles to capture object motion due to the limited diversity of the initial interactions. To overcome this, we introduce an intrinsically motivated reward signal based on world model’s prediction error, which drives a policy to collect more informative trajectories. This results in up to three times more object displacement than random actions, significantly enriching the dataset. Fine-tuning both the vision and world model on these data improves prediction and reconstruction performance. We validate our method in a simulated robotic environment with diverse objects, demonstrating that meaningful visual and physical representations can emerge entirely from self-supervised interaction. This highlights the potential of intrinsically motivated, object-centric learning for autonomous world perception and modeling. |
Katerina Papadimitriou, Gerasimos Potamianos Conformer-Based Multi-Modal Learning for Cued Speech Recognition from Videos Conference IEEE European Signal Processing Conference (EUSIPCO), 2025. @conference{nokey,
title = {Conformer-Based Multi-Modal Learning for Cued Speech Recognition from Videos},
author = {Katerina Papadimitriou, Gerasimos Potamianos},
url = {https://ieeexplore.ieee.org/document/11226352},
doi = {10.23919/EUSIPCO63237.2025.11226352},
year = {2025},
date = {2025-09-08},
urldate = {2025-09-08},
publisher = {IEEE European Signal Processing Conference (EUSIPCO)},
abstract = {This paper presents a novel multimodal framework for automatic cued speech recognition (ACSR) based on a Conformer architecture that directly processes whole upper-body video frames, eliminating the need for explicit segmentation or synchronization of hand and lip regions. Unlike prior approaches that separately process hand and mouth cues, our model jointly learns appearance and skeletal representations, effectively handling hand-lip asynchrony. Our framework consists of two branches (streams): (i) an appearance-based stream, utilizing the ResNet18 model for RGB feature extraction, and (ii) a skeletal-based stream, employing a modulated graph convolutional network (GCN) to process 3D joint coordinates extracted via MediaPipe. To learn temporal dependencies, we integrate a temporal convolutional network (TCN) for short-range temporal modeling and a Conformer encoder for long-range sequence learning. In addition, to enhance feature alignment and improve phoneme sequence learning, we incorporate an auxiliary loss, ensuring robust multimodal fusion. Extensive evaluations on three benchmark cued speech datasets, including French, British English, and Mandarin Chinese, demonstrate that our model achieves state-of-the-art performance, outperforming existing approaches.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
This paper presents a novel multimodal framework for automatic cued speech recognition (ACSR) based on a Conformer architecture that directly processes whole upper-body video frames, eliminating the need for explicit segmentation or synchronization of hand and lip regions. Unlike prior approaches that separately process hand and mouth cues, our model jointly learns appearance and skeletal representations, effectively handling hand-lip asynchrony. Our framework consists of two branches (streams): (i) an appearance-based stream, utilizing the ResNet18 model for RGB feature extraction, and (ii) a skeletal-based stream, employing a modulated graph convolutional network (GCN) to process 3D joint coordinates extracted via MediaPipe. To learn temporal dependencies, we integrate a temporal convolutional network (TCN) for short-range temporal modeling and a Conformer encoder for long-range sequence learning. In addition, to enhance feature alignment and improve phoneme sequence learning, we incorporate an auxiliary loss, ensuring robust multimodal fusion. Extensive evaluations on three benchmark cued speech datasets, including French, British English, and Mandarin Chinese, demonstrate that our model achieves state-of-the-art performance, outperforming existing approaches. |
Stelios Zarifis, Ioannis Kordonis, Petros Maragos Diffusion-Based Forecasting for Uncertainty-Aware Model Predictive Control Conference IEEE European Signal Processing Conference (EUSIPCO), 2025. @conference{nokey,
title = {Diffusion-Based Forecasting for Uncertainty-Aware Model Predictive Control},
author = {Stelios Zarifis, Ioannis Kordonis, Petros Maragos},
url = {https://eurasip.org/Proceedings/Eusipco/Eusipco2025/pdfs/0001877.pdf},
doi = {10.23919/EUSIPCO63237.2025.11226479},
year = {2025},
date = {2025-09-08},
urldate = {2025-09-08},
publisher = {IEEE European Signal Processing Conference (EUSIPCO)},
abstract = {We propose Diffusion-Informed Model Predictive Control (D-I MPC), a generic framework for uncertainty-aware prediction and decision-making in partially observable stochastic systems by integrating diffusion-based time series forecasting models in Model Predictive Control algorithms. In our approach, a diffusion-based time series forecasting model probabilistically estimates the evolution of the system’s stochastic components. These forecasts are then incorporated into MPC algorithms to estimate future trajectories and optimize action selection under the uncertainty of the future. We evaluate the framework on the task of energy arbitrage, where a Battery Energy Storage System participates in the day-ahead electricity market of the New York state. Experimental results indicate that our model-based approach with a diffusion-based forecaster significantly outperforms both implementations with classical forecasting methods and model-free reinforcement learning baselines.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
We propose Diffusion-Informed Model Predictive Control (D-I MPC), a generic framework for uncertainty-aware prediction and decision-making in partially observable stochastic systems by integrating diffusion-based time series forecasting models in Model Predictive Control algorithms. In our approach, a diffusion-based time series forecasting model probabilistically estimates the evolution of the system’s stochastic components. These forecasts are then incorporated into MPC algorithms to estimate future trajectories and optimize action selection under the uncertainty of the future. We evaluate the framework on the task of energy arbitrage, where a Battery Energy Storage System participates in the day-ahead electricity market of the New York state. Experimental results indicate that our model-based approach with a diffusion-based forecaster significantly outperforms both implementations with classical forecasting methods and model-free reinforcement learning baselines. |
Konstantinos Fotopoulos, Christos Garoufis, Petros Maragos Sparse Hybrid Linear-Morphological Networks Conference IEEE European Signal Processing Conference (EUSIPCO), 2025. @conference{nokey,
title = {Sparse Hybrid Linear-Morphological Networks},
author = {Konstantinos Fotopoulos, Christos Garoufis, Petros Maragos},
url = {https://ieeexplore.ieee.org/document/11226484},
doi = {10.23919/EUSIPCO63237.2025.11226484},
year = {2025},
date = {2025-09-08},
urldate = {2025-09-08},
publisher = {IEEE European Signal Processing Conference (EUSIPCO)},
abstract = {We investigate hybrid linear-morphological networks. Recent studies highlight the inherent affinity of morphological layers to pruning, but also their difficulty in training. We propose a hybrid network structure, wherein morphological layers are inserted between the linear layers of the network, in place of activation functions. We experiment with the following morphological layers: 1) maxout pooling layers (as a special case of a morphological layer), 2) fully connected dense morphological layers, and 3) a novel, sparsely initialized variant of (2). We conduct experiments on the Magna-Tag-A-Tune (music autotagging) and CIFAR-10 (image classification) datasets, replacing the linear classification heads of state-of-the-art convolutional network architectures with our proposed network structure for the various morphological layers. We demonstrate that these networks induce sparsity to their linear layers, making them more prunable under L1 unstructured pruning. We also show that on MTAT our proposed sparsely initialized layer achieves slightly better performance than ReLU, maxout, and densely initialized max-plus layers, and exhibits faster initial convergence.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
We investigate hybrid linear-morphological networks. Recent studies highlight the inherent affinity of morphological layers to pruning, but also their difficulty in training. We propose a hybrid network structure, wherein morphological layers are inserted between the linear layers of the network, in place of activation functions. We experiment with the following morphological layers: 1) maxout pooling layers (as a special case of a morphological layer), 2) fully connected dense morphological layers, and 3) a novel, sparsely initialized variant of (2). We conduct experiments on the Magna-Tag-A-Tune (music autotagging) and CIFAR-10 (image classification) datasets, replacing the linear classification heads of state-of-the-art convolutional network architectures with our proposed network structure for the various morphological layers. We demonstrate that these networks induce sparsity to their linear layers, making them more prunable under L1 unstructured pruning. We also show that on MTAT our proposed sparsely initialized layer achieves slightly better performance than ReLU, maxout, and densely initialized max-plus layers, and exhibits faster initial convergence. |
Katerina Papadimitriou, Gerasimos Potamianos A Multi-Stream Framework Utilizing 3D Human Reconstruction for Cued Speech Recognition Conference Interspeech, 2025. @conference{nokey,
title = {A Multi-Stream Framework Utilizing 3D Human Reconstruction for Cued Speech Recognition},
author = {Katerina Papadimitriou, Gerasimos Potamianos},
url = {https://www.isca-archive.org/interspeech_2025/papadimitriou25_interspeech.pdf},
doi = {10.21437/Interspeech.2025-2195},
year = {2025},
date = {2025-08-17},
urldate = {2025-08-17},
publisher = {Interspeech},
abstract = {In this paper, we propose a novel multi-stream framework for automatic cued speech recognition (ACSR) that directly processes the upper-body region, addressing hand-lip asynchrony without requiring explicit segmentation or synchronization. Our model integrates two distinct modalities: (i) an appearance based stream leveraging the ResNet18 for feature extraction and (ii) a skeletal stream based on a modulated graph convolutional network (GCN). For graph construction, we incorporate, for the first time in ACSR, 3D pose parameters inferred from the PIXIE model. Both modalities are coupled with temporal convolution for short-range dynamics learning and a BiGRU encoder for long-term sequence modeling. In addition, we introduce an alignment module that combines CTC with two auxiliary losses, improving each modality performance and enabling effective late fusion during inference. Our model achieves state-of-the-art performance across three benchmark datasets, demonstrating its effectiveness.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
In this paper, we propose a novel multi-stream framework for automatic cued speech recognition (ACSR) that directly processes the upper-body region, addressing hand-lip asynchrony without requiring explicit segmentation or synchronization. Our model integrates two distinct modalities: (i) an appearance based stream leveraging the ResNet18 for feature extraction and (ii) a skeletal stream based on a modulated graph convolutional network (GCN). For graph construction, we incorporate, for the first time in ACSR, 3D pose parameters inferred from the PIXIE model. Both modalities are coupled with temporal convolution for short-range dynamics learning and a BiGRU encoder for long-term sequence modeling. In addition, we introduce an alignment module that combines CTC with two auxiliary losses, improving each modality performance and enabling effective late fusion during inference. Our model achieves state-of-the-art performance across three benchmark datasets, demonstrating its effectiveness. |
Dimitris N. Makropoulos, Panagiotis P. Filntisis, Aristides Prospathopoulos, Dimitris Kassis, Antigoni Tsiami, Petros Maragos Improving Classification of Marine Mammal Vocalizations Using Vision Transformers and Phase-Related Features Conference IEEE International Conference on Digital Signal Processing (DSP), 2025. @conference{nokey,
title = {Improving Classification of Marine Mammal Vocalizations Using Vision Transformers and Phase-Related Features},
author = {Dimitris N. Makropoulos, Panagiotis P. Filntisis, Aristides Prospathopoulos, Dimitris Kassis, Antigoni Tsiami, Petros Maragos},
url = {https://ieeexplore.ieee.org/document/11075076},
doi = {10.1109/DSP65409.2025.11075076},
year = {2025},
date = {2025-06-25},
urldate = {2025-06-25},
publisher = {IEEE International Conference on Digital Signal Processing (DSP)},
abstract = {In this study, we investigate the relative performance of Vision Transformers (ViTs) compared to convolution-based neural networks in categorizing vocalizations from a medium-sized marine mammal dataset. Additionally, we evaluate whether phase information derived from Fourier decomposition can serve as a complementary source of useful information to magnitude for classification tasks. Our study focuses on bioacoustics, utilizing the publicly available Watkins Marine Mammal Sound Database, which contains sound clips identified as originating from 32 marine mammal species. In this framework, we first trained convolution-based networks (ResNet-101, MobileNetV3) and Transformer-based networks (ViT B-16, Swin Transformer V2) on log-magnitude spectrograms (baseline models). In a second set of experiments, we incorporated the derivative of unwrapped phase from the Fourier representation into the magnitude spectrograms. Our results show that (a) Shifted Window (Swin) Transformers outperform MobileNets and achieve performance similar to ResNets while maintaining lower computational complexity and (b) the inclusion of phase derivatives into spectrograms leads to (i) consistently improved performance metrics across all biosignal categories for Swin Transformers and (ii) enhanced classification ability for both convolution-based and self-attention-based networks, particularly for the narrow-band frequency modulated (FM) whistles emitted by delphinids,},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
In this study, we investigate the relative performance of Vision Transformers (ViTs) compared to convolution-based neural networks in categorizing vocalizations from a medium-sized marine mammal dataset. Additionally, we evaluate whether phase information derived from Fourier decomposition can serve as a complementary source of useful information to magnitude for classification tasks. Our study focuses on bioacoustics, utilizing the publicly available Watkins Marine Mammal Sound Database, which contains sound clips identified as originating from 32 marine mammal species. In this framework, we first trained convolution-based networks (ResNet-101, MobileNetV3) and Transformer-based networks (ViT B-16, Swin Transformer V2) on log-magnitude spectrograms (baseline models). In a second set of experiments, we incorporated the derivative of unwrapped phase from the Fourier representation into the magnitude spectrograms. Our results show that (a) Shifted Window (Swin) Transformers outperform MobileNets and achieve performance similar to ResNets while maintaining lower computational complexity and (b) the inclusion of phase derivatives into spectrograms leads to (i) consistently improved performance metrics across all biosignal categories for Swin Transformers and (ii) enhanced classification ability for both convolution-based and self-attention-based networks, particularly for the narrow-band frequency modulated (FM) whistles emitted by delphinids, |
Chrysa Pratikaki, Panagiotis Filntisis, Athanasios Katsamanis, Anastasios Roussos, Petros Maragos A Transformer-Based Framework for Greek Sign Language Production using Extended Skeletal Motion Representations Conference ACM ACM International Conference on PErvasive Technologies Related to Assistive Environments (PETRA), 2025. @conference{nokey,
title = {A Transformer-Based Framework for Greek Sign Language Production using Extended Skeletal Motion Representations},
author = {Chrysa Pratikaki, Panagiotis Filntisis, Athanasios Katsamanis, Anastasios Roussos, Petros Maragos},
url = {https://dl.acm.org/doi/epdf/10.1145/3733155.3733204},
doi = {10.1145/3733155.3733204},
year = {2025},
date = {2025-06-25},
urldate = {2025-06-25},
publisher = {ACM International Conference on PErvasive Technologies Related to Assistive Environments (PETRA)},
organization = {ACM},
abstract = {Sign Languages are the primary form of communication for Deaf communities across the world. To break the communication barriers between the Deaf and Hard-of-Hearing and the hearing communities, it is imperative to build systems capable of translating the spoken language into sign language and vice versa. Building on insights from previous research, we propose a deep learning model for Sign Language Production (SLP), which to our knowledge is the first attempt on Greek SLP. We tackle this task by utilizing a transformer-based architecture that enables the translation from text input to human pose keypoints, and the opposite. We evaluate the effectiveness of the proposed pipeline on the Greek SL dataset Elementary23, through a series of comparative analyses and ablation studies. Our pipeline’s components, which include data-driven gloss generation, training through video to text translation and a scheduling algorithm for teacher forcing - auto-regressive decoding seem to actively enhance the quality of produced SL videos.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
Sign Languages are the primary form of communication for Deaf communities across the world. To break the communication barriers between the Deaf and Hard-of-Hearing and the hearing communities, it is imperative to build systems capable of translating the spoken language into sign language and vice versa. Building on insights from previous research, we propose a deep learning model for Sign Language Production (SLP), which to our knowledge is the first attempt on Greek SLP. We tackle this task by utilizing a transformer-based architecture that enables the translation from text input to human pose keypoints, and the opposite. We evaluate the effectiveness of the proposed pipeline on the Greek SL dataset Elementary23, through a series of comparative analyses and ablation studies. Our pipeline’s components, which include data-driven gloss generation, training through video to text translation and a scheduling algorithm for teacher forcing - auto-regressive decoding seem to actively enhance the quality of produced SL videos. |
Artemis Androni, Christos Garoufis, Athanasia Zlatintsi, Petros Maragos Feature-Level Multimodal Fusion for Relapse Detection in Patients with Psychosis Conference IEEE IEEE International Conference on Digital Signal Processing (DSP), 2025. @conference{nokey,
title = {Feature-Level Multimodal Fusion for Relapse Detection in Patients with Psychosis},
author = {Artemis Androni, Christos Garoufis, Athanasia Zlatintsi, Petros Maragos},
url = {https://ieeexplore.ieee.org/document/11075191},
doi = {10.1109/DSP65409.2025.11075191},
year = {2025},
date = {2025-06-25},
urldate = {2025-06-25},
publisher = {IEEE International Conference on Digital Signal Processing (DSP)},
organization = {IEEE},
abstract = {In this work, we aim to improve the detection of relapses in patients with psychotic disorders (i.e., bipolar disorder and schizophrenia) using spontaneous speech data from patient-clinician interviews and physiological signals from wearable sensors. To achieve this, we propose a joint autoencoder framework, trained with coupled data from both modalities, projecting them through distinct encoder branches into a unified latent space and then separately decoding them. We experiment with convolutional and LSTM-based autoencoders for the speech data, whereas we adapt convolutional autoencoders for the physiological signals. Our experimental results show that the proposed multimodal fusion scheme consistently outperforms unimodal baselines, with the introduced LSTM-based autoencoders proving to be an effective alternative to convolutional ones for relapse detection from speech. Moreover, through ablation experiments, we confirm that not only both modalities contribute positively to the joint framework, but our approach outperforms unimodally trained baselines when only the respective modality is available.},
keywords = {},
pubstate = {published},
tppubtype = {conference}
}
In this work, we aim to improve the detection of relapses in patients with psychotic disorders (i.e., bipolar disorder and schizophrenia) using spontaneous speech data from patient-clinician interviews and physiological signals from wearable sensors. To achieve this, we propose a joint autoencoder framework, trained with coupled data from both modalities, projecting them through distinct encoder branches into a unified latent space and then separately decoding them. We experiment with convolutional and LSTM-based autoencoders for the speech data, whereas we adapt convolutional autoencoders for the physiological signals. Our experimental results show that the proposed multimodal fusion scheme consistently outperforms unimodal baselines, with the introduced LSTM-based autoencoders proving to be an effective alternative to convolutional ones for relapse detection from speech. Moreover, through ablation experiments, we confirm that not only both modalities contribute positively to the joint framework, but our approach outperforms unimodally trained baselines when only the respective modality is available. |
