Research map / Architectures and efficiency / Attention and transformers
Design and Analysis of Efficient Attention in Transformers for Social Group Activity Recognition
Masato Tamura
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 |
|---|---|---|---|
| decompose group queries into location and shared layout queries (ablation) | Ours · Volleyball group activity recognition | accuracy: 95.1 -> 95.4 | Improved |
| decompose group queries into location and shared layout queries (ablation) | Ours · Volleyball social group activity recognition | accuracy: 63.7 -> 64.4 | Improved |
| add intra-group attention after inter-group attention (ablation) | Ours · Volleyball group activity recognition | accuracy: 94.1 -> 95.4 | Improved |
| add intra-group attention after inter-group attention (ablation) | Ours · Volleyball social group activity recognition | accuracy: 60.3 -> 64.4 | Improved |
| apply inter-group before intra-group attention instead of reverse order (ablation) | Ours · Volleyball group activity recognition | accuracy: 94.5 -> 95.4 | Improved |
| apply inter-group before intra-group attention instead of reverse order (ablation) | Ours · Volleyball social group activity recognition | accuracy: 63.2 -> 64.4 | Improved |
| use divided attention with 12 embeddings instead of naive attention with 6 (ablation) | Ours · Volleyball group activity recognition | accuracy: 95.7 -> 95.4 | Hurt |
| use divided attention with 12 embeddings instead of naive attention with 6 (ablation) | Ours · Volleyball social group activity recognition | accuracy: 61.6 -> 64.4 | Improved |
| use divided attention instead of naive attention on member identification (ablation) | Ours · Volleyball group member identification | duplicated ratio (lower is better): 11.7 -> 9.94 | Improved |
| add joint individual action recognition training (ablation) | Ours · Volleyball social group activity recognition | accuracy: 62.1 -> 64.4 | Improved |
| add joint individual action recognition training | Ehsanpour et al. · Volleyball social group activity recognition | accuracy: 57.6 -> 44.5 | Hurt |
| use multiple embeddings per group (Ours) instead of one embedding (previous method) | Ours · Volleyball social group activity recognition | accuracy: 60.6 -> 64.4 | Improved |
Closest papers by meaning
- Hunting Group Clues with Transformers for Social Group Activity Recognition (ECCV 2022 · similarity 0.94)
- GroupFormer: Group Activity Recognition With Clustered Spatial-Temporal Transformer (ICCV 2021 · similarity 0.91)
- Joint Learning of Social Groups, Individuals Action and Sub-group Activities in Videos (ECCV 2020 · similarity 0.89)
- Detector-Free Weakly Supervised Group Activity Recognition (CVPR 2022 · similarity 0.88)
- Actor-Transformers for Group Activity Recognition (CVPR 2020 · similarity 0.88)
- Bi-Causal: Group Activity Recognition via Bidirectional Causality (CVPR 2024 · similarity 0.84)
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.
