Research map / Label-efficient and robust learning / Recognition and long-tail
Hunting Group Clues with Transformers for Social Group Activity Recognition
Masato Tamura, Rahul Vishwakarma, Ravigopal Vennelakanti
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| sort group member points by X coordinate rather than Y (ablation) | Ours · Volleyball social group activity | accuracy: 55.5 -> 60.6 | Improved |
| add group annotations to training (ablation) | Ours · Volleyball group activity | accuracy: 95.0 -> 96.0 | Improved |
| add group annotations to training (ablation) | GroupFormer · Volleyball group activity (detection-based) | accuracy: 95.0 -> 93.2 | Hurt |
| add group annotations to training (ablation) | GroupFormer · Volleyball group activity (ground-truth boxes) | accuracy: 95.7 -> 96.1 | Improved |
| replace region features with transformer-generated group features (ours vs GroupFormer) (ablation) | Ours · Volleyball social group activity | accuracy: 48.8 -> 60.6 | Improved |
| replace region features with transformer-generated group features (ours vs Ehsanpour et al.) (ablation) | Ours · Volleyball social group activity | accuracy: 44.5 -> 60.6 | Improved |
| use our method instead of Ehsanpour et al. for Collective Activity social groups (ablation) | Ours · Collective Activity social group activity | mAP: 51.3 -> 46.0 | Hurt |
| use predicted rather than ground-truth boxes (ablation) | GroupFormer · Volleyball group activity | accuracy: 95.7 -> 95.0 | Hurt |
| use predicted rather than ground-truth boxes (ablation) | GroupFormer · Collective Activity group activity | accuracy: 96.3 -> 85.2 | Hurt |
| rely on learned group size with few training samples (size 4 on Collective Activity) (ablation) | Ours · Collective Activity Talking | mAP: 6.56 | Hurt |
| larger training data ratio for group size (ablation) | Ours · Volleyball group size analysis | accuracy: 83.6 at 36% vs 26.1 at 6% | Improved |
| our method vs Ehsanpour et al. on group size 3 | Ours · Volleyball group size 3 | accuracy: 61.2 -> 52.4 | Hurt |
Closest papers by meaning
- Design and Analysis of Efficient Attention in Transformers for Social Group Activity Recognition (IJCV 2024 · similarity 0.94)
- GroupFormer: Group Activity Recognition With Clustered Spatial-Temporal Transformer (ICCV 2021 · similarity 0.92)
- Joint Learning of Social Groups, Individuals Action and Sub-group Activities in Videos (ECCV 2020 · similarity 0.92)
- Detector-Free Weakly Supervised Group Activity Recognition (CVPR 2022 · similarity 0.90)
- Learning Group Activity Features Through Person Attribute Prediction (CVPR 2024 · similarity 0.86)
- Bi-Causal: Group Activity Recognition via Bidirectional Causality (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
