Reproduce “Temporal Feature Distillation for Label-Efficient Precise Event Spotting in Sports Videos” 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

Hao Xu, Xinyu Wei, Sam Wells, Sunil Aryal

A semi-supervised framework named Temporal Feature Distillation (TFD) improves precise event spotting in sports videos by integrating a Transformer Gate Shift (TGS) module for temporal modeling, Temporal Motion Augmentation (TMA) for motion-focused views, and a feature-level distillation objective. This method achieves superior performance across various sports datasets, yielding over 4.5 mAP improvement on FSPerf with 10% labeled data and matching 100% supervised baselines on two datasets with 80% labels.

Key insights

  • Directly aligning backbone features in a self-distillation framework (TFD) is crucial for preserving temporal boundary cues required for precise event spotting, as projection-level invariance used in DINO-style methods suppresses this fine-grained information.
  • Integrating a Transformer Gate Shift (TGS) module effectively enables 2D Vision Transformers to capture multi-scale temporal information without computationally expensive 3D architectures, enhancing their capability for fine-grained video analysis.
  • Tailored data augmentation that emphasizes motion-sensitive regions (TMA) and a supervised warm-up followed by a distillation ramp-up schedule are essential for learning robust, localized, and boundary-sensitive representations in semi-supervised PES.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.