Reproduce “Group-DINOmics: Incorporating People Dynamics into DINO for Self-supervised Group Activity Feature Learning” 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

Ryuki Tezuka, Chihiro Nakatani, Norimichi Ukita

Group-DINOmics introduces a self-supervised method for learning comprehensive group activity features by augmenting DINOv3 with pretext tasks focused on person dynamics and group-relevant object context. This approach achieves substantial improvements in group activity retrieval (e.g., 21.6% Hit@1 gain on VBD) and competitive performance in supervised recognition, without requiring manual activity annotations.

Key insights

  • Sequentially training dynamics-aware tasks before global context-aware tasks proves more effective in learning comprehensive GAFs.
  • Inpainting group-relevant objects forces the model to rely on global contextual features to infer object locations, enhancing the GAF's global context awareness.
  • Auxiliary tasks, which optimize the DINOv3 backbone directly for dynamics and object localization, are crucial for embedding activity-relevant information into frame-level features.

Summary by alphaXiv.

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