Reproduce “Efficient tracking of team sport players with few game-specific annotations” 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
Adrien Maglo, Astrid Orcesi, Quoc-Cuong Pham
A new method enables efficient tracking and identification of team sport players with only a few game-specific annotations, leveraging an incremental learning framework and releasing a public rugby sevens dataset. It achieves improved accuracy in challenging environments like rugby, where players have similar appearances and frequent occlusions.
Key insights
- Robust player re-identification can be achieved with a minimal number of annotations (approximately 3.5 per player), significantly reducing the annotation burden.
- An incremental training strategy, initially freezing the single-image re-ID network, enables fast and interactive model refinement, which can then be followed by full network training for optimal performance.
- Tailored re-identification models with specific association algorithms are necessary for team sports, as generic MOT methods struggle with inherent challenges like appearance similarity and frequent occlusions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.