3D Human Shape and Pose Estimation from Multi-view Images

3D human shape and pose estimation is at the core of this thesis, which presents an end-to-end pipeline for reconstructing accurate body models from multi-view images of fitness exercises. The process begins with videos from four cameras, from which a pre-trained model extracts 2D keypoints for each single frame. These keypoints are then fused by a specialized neural network to reconstruct precise 3D positions, overcoming the limitations of monocular methods. In a subsequent step, a regression model computes SMPL-X parameters to provide a detailed representation that captures both the overall pose and intricate body features. Overall, integrating 2D detection, multi-view 3D reconstruction, and SMPL-X modeling establishes a solid foundation for innovative systems in human motion analysis and optimization, with promising applications in virtual coaching and fitness.