sst |
Ray-based Multi-view 2D to 3D pose Lifter by Seyed Abolfazl GHASEMZADEH -
Vendredi 20 juin 2025 à 16h30 - Auditoire BARB94 - Place Sainte Barbe - 1348 Louvain-la-Neuve -
Accurately estimating 3D human poses from 2D images remains a core challenge in computer vision, particularly in unconstrained, real-world environments. Traditional multi-view methods depend heavily on datasets with paired images and 3D annotations, which are expensive to collect and typically limited to controlled laboratory settings.
This work introduces the Ray-based Multi-view 3D Pose Lifter (RMPL), a novel transformer-based architecture that reconstructs 3D poses from multi-view 2D keypoints—entirely in the pose space, without requiring camera calibration or image-based training data. Leveraging a synthetic data pipeline that mimics the noise of real 2D detectors, RMPL is both generalizable and scalable across diverse scenarios.
The proposed framework achieves strong results on standard benchmarks and adapts easily to new camera setups and 2D pose detectors. By removing the need for costly 3D datasets, it paves the way for broader, more practical deployment of 3D human pose estimation in sports analytics, medical diagnostics, virtual reality, and robotics.
Jury members :
Prof. Christophe De Vleeschouwer (UCLouvain)(Promoteur)
Prof. Laurent Francis (UCLouvain) (Président)
Prof. Benoît Macq (UCLouvain) (Secrétaire)
Prof. Vincent François-Lavet (VUAmsterdam)
Prof. Anthony Cioppa (ULiege)
Pay attention : the pubic defense of Seyed Abolfazl GHASEMZADEH will also take place in the form of a videoconference
