Research map / Label-efficient and robust learning / Recognition and long-tail
Revisiting Skeleton-Based Action Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Represent poses as 3D heatmap volumes and use a 3D-CNN instead of graph-coordinate processing | Skeleton-based action recognition · NTU60-XSub | Top-1 accuracy: 91.5 -> 93.7 | Helped |
| Add limb heatmaps and ensemble their predictions with joint heatmaps (ablation) | PoseConv3D · NTU60-XSub | Top-1 accuracy: 93.7 -> 94.1 | Helped |
| Add subjects-centered cropping (ablation) | PoseConv3D · FineGYM | Mean Top-1 accuracy: 91.7% to 92.7% | Helped |
| Sample frames uniformly across the full clip rather than at fixed temporal strides (ablation) | PoseConv3D · FineGYM and NTU-60 | Top-1 accuracy | Helped |
| Use bidirectional rather than unidirectional RGB-pose lateral connections (ablation) | RGBPose-Conv3D · FineGYM | Top-1 accuracy: 93.4 -> 93.6 (1-clip); 93.8 -> 94.1 (10-clip) | Helped |
| Add early RGB-pose fusion to late fusion (ablation) | RGBPose-Conv3D · FineGYM and NTU-60 | Top-1 accuracy: 92.6 / 93.4 -> 93.6 / 94.1; 95.5 / 96.0 -> 96.2 / 96.5 | Helped |
| Increase input resolution or double network width (ablation) | Pose-SlowOnly · NTU60-XSub | Top-1 accuracy: 93.7 -> 93.6 (HR); 93.7 -> 93.7 (wd) | No clear effect |
| Compress heatmaps into coordinate triplets and reconstruct pseudo heatmaps (ablation) | PoseConv3D with HRNet poses · FineGYM | Mean Top-1 accuracy: 93.6 -> 93.2 | Hurt |
| Use coordinate triplets rather than heatmap inputs (ablation) | PoseConv3D with MobileNet poses · FineGYM | Mean Top-1 accuracy: 92.7 -> 90.7 | Hurt |
| Use estimated 2D poses rather than estimated or sensor-collected 3D poses (ablation) | Skeleton action recognition · FineGYM | Mean Top-1 accuracy: 87.0 -> 92.0 (VIBE-3D to HRNet-2D) | Helped |
| Lift 2D poses into 3D before recognition (ablation) | Skeleton action recognition · FineGYM | Mean Top-1 accuracy: 92.0 -> 90.0 (FrameLift); 92.0 -> 90.2 (VideoLift) | Hurt |
| Use a single accumulated heatmap volume for all people rather than separate person representations (ablation) | PoseConv3D · Volleyball | Top-1 accuracy: 75.3%; 85.7%; 87.9% versus 91.3% | Helped |
Closest papers by meaning
- InfoGCN: Representation Learning for Human Skeleton-Based Action Recognition (CVPR 2022 · similarity 0.89)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.88)
- Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.88)
- Skeleton-based Action Recognition with Non-linear Dependency Modeling and Hilbert-Schmidt Independence Criterion (AAAI 2025 · similarity 0.87)
- Recovering Complete Actions for Cross-dataset Skeleton Action Recognition (NeurIPS 2024 · similarity 0.87)
- Dynamic Semantic-Based Spatial Graph Convolution Network for Skeleton-Based Human Action Recognition (AAAI 2024 · 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.