Automated boxing performance analysis presents significant challenges due to the scarcity of high-quality annotated datasets, the complexity of body-to-body interactions, the variability of postures, and frequent occlusions in real-world footage. Manual annotation is time-consuming, error-prone, and difficult to scale, motivating the use of synthetic data generation as an alternative. In this thesis, we present a modular and reproducible pipeline for creating photorealistic 3D boxing scenes, leveraging parametric human models (SMPL), motion capture-driven animations, and procedurally generated environments representing diverse contexts (e.g., training gym, Olympic competition, Las Vegas events). The pipeline incorporates a multi-level automatic annotation system, supported by manual verification, to capture detailed information on actions, biomechanical postures, and outcome categories. A dedicated web-based annotation tool offers synchronized multi-view playback, temporal segmentation, and hierarchical labeling. The rendering engine generates large-scale datasets with rich variations in camera angles, lighting conditions, and body morphologies, enabling robust training samples. Experimental evaluations demonstrate that deep learning models trained on these synthetic data achieve competitive accuracy in action recognition and pose estimation, even when applied to real footage. This work highlights the potential of synthetic datasets in advancing computer vision systems for combat sports performance analysis and AI-assisted biomechanical modeling.
| Date | 12 Dec 2025 |
|---|
| Original language | French |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Sheldon Andrews (Supervisor) & David Labbé (Co-supervisor) |
|---|
Takouti, Y. (Author),
Andrews (Supervisor) &
Labbé (Co-supervisor),
12 Dec 2025Student thesis: Master's thesis › Master in Engineering: Information Technology Engineering