Research map / Graph learning / Knowledge and temporal graphs
Representing Videos As Discriminative Sub-Graphs for Action Recognition
Dong Li, Zhaofan Qiu, Yingwei Pan, Ting Yao, Houqiang Li, Tao Mei
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 |
|---|---|---|---|
| Replace clip-level classification with discriminative sub-graph reasoning (ablation) | LGD Base · Something-Something V2 val | top-1 accuracy: 60.8 -> 63.2 | Improved |
| Use 4-node rather than 3-node sub-graphs (ablation) | Base+Sub-graph · Something-Something V2 val | top-1 accuracy: 62.7 -> 63.2 | Improved |
| Use 5-node rather than 4-node sub-graphs (ablation) | Base+Sub-graph · Something-Something V2 val | top-1 accuracy: 63.2 -> 62.2 | Hurt |
| Add tubelet coordinate features (ablation) | Base+Sub-graph (4 nodes) · Something-Something V2 val | top-1 accuracy: +0.3% | Improved |
| Add relative coordinate changes to graph edges (ablation) | Node Coord (4 nodes) · Something-Something V2 val | top-1 accuracy: +0.6% | Improved |
| Combine sub-graphs across three scales (ablation) | Edge Coord (4 nodes) · Something-Something V2 val | top-1 accuracy: 64.1 -> 65.0 | Improved |
| Use sub-graph prototypes instead of whole-graph GCN reasoning (ablation) | Base+GCN · Something-Something V2 val | top-1 accuracy: 61.3 -> 63.2 | Improved |
| Use MUSLE instead of TEA | ResNet-50 action recognition · Something-Something V1 val | top-1 accuracy: 0.6% | Improved |
| Use MUSLE instead of STM | ResNet-50 action recognition · Something-Something V2 val | top-1 accuracy: 0.8% | Improved |
| Transfer Kinetics-trained action prototypes without fine-tuning | MUSLE · UCF101 | accuracy: 9.6% | Improved |
| Transfer Kinetics-trained action prototypes without fine-tuning | MUSLE · HMDB51 | accuracy: 13.5% | Improved |
Closest papers by meaning
- G-TAD: Sub-Graph Localization for Temporal Action Detection (CVPR 2020 · similarity 0.89)
- Unsupervised Action Discovery and Localization in Videos (ICCV 2017 · similarity 0.88)
- Learning Asynchronous and Sparse Human-Object Interaction in Videos (CVPR 2021 · similarity 0.88)
- Action Genome: Actions As Compositions of Spatio-Temporal Scene Graphs (CVPR 2020 · similarity 0.88)
- Unified Graph Structured Models for Video Understanding (ICCV 2021 · similarity 0.87)
- Temporally-Weighted Hierarchical Clustering for Unsupervised Action Segmentation (CVPR 2021 · similarity 0.87)
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.
