MSc Thesis · 2025 – 2026 · Grade 9/10
“From Pixels to Poses”
Monocular 3D Human Pose Estimation in Broadcast Football
Vrije Universiteit Amsterdam — in collaboration with Borussia Dortmund (BVB) and Tactive Sports Amsterdam
56.3%MPJPE reduction
1invited publication
- Built a five-stage pipeline that segments, tracks and re-identifies individual players in broadcast video and reconstructs their 3D pose — SAM 3, ByteTrack with OSNet re-ID, ViTPose-H and MotionBERT assembled into one system.
- Designed and trained a Spatio-Temporal Pose Refinement Transformer as a post-hoc correction stage, with a learnable skeleton adjacency bias and a composite loss enforcing temporal smoothness, bone-length integrity and bilateral symmetry.
- Cut mean per joint position error from 184.0 mm to 80.4 mm, and mean body facing direction error from 52.3° to 21.5°.
- Established that the gain comes from learned structural priors, not filtering: Gaussian temporal smoothing improves the baseline by less than 1%.
- Generated ground truth from a 90-minute match in a modified Unity football simulator, and trained on the Snellius national HPC cluster.
Publication — H. Joël. From Pixels to Poses: A Supervised Spatio-Temporal Refinement Pipeline for Monocular 3D Human Pose Estimation in Broadcast Football. Invited for journal publication; manuscript in preparation, 2026.