Research map / Video / Video-language understanding
COMPOSER: Compositional Reasoning of Group Activity in Videos with Keypoint-Only Modality
Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, Hans Peter Graf
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 |
|---|---|---|---|
| use keypoint-only input instead of RGB features (ablation) | COMPOSER RGB variant · Volleyball Olympic split | group activity accuracy (%): 81.1 -> 95.1 | Improved |
| add contrastive clustering for scale agreement (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 93.4 -> 94.6 | Improved |
| use contrastive clustering instead of label consistency (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 93.9 -> 94.6 | Improved |
| increase number of scales from 1 to 4 (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 91.2 -> 94.6 | Improved |
| add auxiliary predictions at each scale and block (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 93.3 -> 94.6 | Improved |
| add Multiscale Transformer instead of a flat model (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 88.1 -> 94.6 | Improved |
| add Actor Dropout augmentation (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 94.0 -> 94.6 | Improved |
| add horizontal flip augmentation (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 94.0 -> 94.6 | Improved |
| add horizontal and vertical move augmentation (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 94.2 -> 94.6 | Improved |
| fine-to-coarse token order instead of reverse (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 86.8 -> 94.6 | Improved |
| separate transformer encoder per scale instead of one encoder for all tokens (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 92.4 -> 94.6 | Improved |
| use fixed person grouping instead of time-varying grouping (ablation) | COMPOSER · Volleyball Original split | group activity accuracy (%): 92.8 -> 94.6 | Improved |
Later papers that compared with this method
1 later accepted paper reports COMPOSER as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SkeleTR: Towards Skeleton-based Action Recognition in the Wild (ICCV 2023 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GroupFormer — GroupFormer: Group Activity Recognition With Clustered Spatial-Temporal Transformer (ICCV 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Detector-Free Weakly Supervised Group Activity Recognition (CVPR 2022 · similarity 0.89)
- Temporal Modular Networks for Retrieving Complex Compositional Activities in Videos (ECCV 2018 · similarity 0.88)
- Context-Interactive Reasoning for Group Activity Detection (ECCV 2026 · similarity 0.88)
- GroupFormer: Group Activity Recognition With Clustered Spatial-Temporal Transformer (ICCV 2021 · similarity 0.87)
- Learning Visual Context for Group Activity Recognition (AAAI 2021 · similarity 0.87)
- VELOCITI: Benchmarking Video-Language Compositional Reasoning with Strict Entailment (CVPR 2025 · 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.
