Reproduce “SMART: SMPLest-X Mesh Adaptation and RAFT Tracking for Soccer Pose Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Parthsarthi Rawat
The SMART pipeline estimates per-player 3D joint positions of soccer players from broadcast video by integrating SMPLest-X, RAFT-small optical flow, and domain-specific adaptations like foot-plane anchoring. This approach achieved a 38.6% improvement over the FIFA competition baseline on the validation set, scoring 0.647.
Key insights
- Domain-adaptive fine-tuning with targeted strategies like stratified data splits, multi-task depth supervision, and broadcast augmentation is crucial for successfully applying general 3D body models to specialized domains.
- Robust camera pose estimation, particularly using dense optical flow with outlier rejection, is a dominant factor in achieving high global pose accuracy, constituting approximately 55% of the total competition score.
- Accurate world-space grounding via foot-plane anchoring significantly reduces both global and local pose errors, demonstrating its importance for placing players correctly on the pitch plane.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.