Research map / Video / Action recognition
SkeleTR: Towards Skeleton-based Action Recognition in the Wild
Haodong Duan, Mingze Xu, Bing Shuai, Davide Modolo, Zhuowen Tu, Joseph Tighe, Alessandro Bergamo
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 |
|---|---|---|---|
| add transformer inter-sequence modeling (ablation) | SkeleTR-S · Posetics | Top-1 accuracy (%): 48.21 -> 52.81 | Improved |
| use transformer encoders instead of non-local block (ablation) | inter-sequence modeling · Posetics | Top-1 accuracy (%): 48.43 -> 52.81 | Improved |
| use Mix Pooling instead of single 1x1 pooling (ablation) | SkeleTR-S · Posetics | Top-1 accuracy (%): 52.81 -> 53.90 | Improved |
| use full T x P feature aggregation instead of Mix Pooling (ablation) | SkeleTR-S · Posetics | Top-1 accuracy (%): 52.78 vs 53.90; 5.211 vs 0.594 GFLOPs | No clear effect |
| adaptive sequence number M instead of fixed (ablation) | SkeleTR-L · Posetics | Top-1 and GFLOPs: 54.95 -> 55.06; 6.653 -> 4.934 GFLOPs | No clear effect |
| use many short sequences (40x15) instead of few long sequences (2x300) (ablation) | SkeleTR · Posetics | Top-1 accuracy (%): +2.1% (SkeleTR-S), +3% (SkeleTR-L) | Improved |
| joint training on Posetics and AVA instead of AVA from scratch (ablation) | SkeleTR-S · AVA v2.2 (10s window) | overall mAP: 21.33 -> 25.69 | Improved |
| pretrain on Posetics then finetune on AVA (ablation) | SkeleTR-S · AVA v2.2 (10s window) | overall mAP: 21.33 -> 23.11 | Improved |
| late-fuse SkeleTR with RGB recognizer | TimeSformer / ACAR-R101 · AVA v2.2 | overall mAP: 31.30 -> 34.34 | Improved |
| use SkeleTR short-sequence design with multi-person modeling instead of ST-GCN++ | Posetics classification · Posetics | Top-1 accuracy (%): 54.1 -> 57.9 | Improved |
| use short sequence format when ground-truth skeleton sequences are available | SkeleTR · tracked skeleton input | accuracy | No clear effect |
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.
- COMPOSER — COMPOSER: Compositional Reasoning of Group Activity in Videos with Keypoint-Only Modality (ECCV 2022) · 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
- Ske2Grid: Skeleton-to-Grid Representation Learning for Action Recognition (ICML 2023 · similarity 0.89)
- SkelHCC: A Hyperbolic CLIP-Driven Cache Adaptation Framework for Skeleton-based One-Shot Action Recognition (ICML 2026 · similarity 0.89)
- Skeleton-based Action Recognition with Non-linear Dependency Modeling and Hilbert-Schmidt Independence Criterion (AAAI 2025 · similarity 0.87)
- Stitch, Contrast, and Segment: Learning a Human Action Segmentation Model Using Trimmed Skeleton Videos (AAAI 2025 · similarity 0.87)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.87)
- Heterogeneous Skeleton-Based Action Representation Learning (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.
