Reproduce “A Hierarchical Deep Temporal Model for Group Activity Recognition” 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
Moustafa Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, Greg Mori
Researchers from Simon Fraser University developed a hierarchical deep temporal model that employs stacked LSTMs to recognize group activities by first modeling individual person actions and then aggregating these for group-level temporal analysis. The model achieved 81.5% accuracy on the Collective Activity Dataset and 51.1% on the newly introduced Volleyball Dataset, outperforming various deep learning baselines.
Key insights
- The first-stage LSTM, modeling individual person-level temporal dynamics, contributes more significantly to overall performance than the second-stage group-level LSTM.
- Aggregating individual person features through max-pooling is an effective strategy for creating a robust scene-level representation for group activity recognition.
- Capturing both hierarchical structure (individual to group) and temporal dynamics provides a more discriminative and robust approach for group activity recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.