Research map / Video / Action recognition
Spatiotemporal Self-Attention Modeling with Temporal Patch Shift for Action Recognition
Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, Xian-Sheng Hua, Lei Zhang
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 temporal patch shift with a single center patch (ablation) | Swin-T · Something-Something V1 | top-1: 40.6 -> 45.3 | Improved |
| shift evenly distributed patches from 3 frames instead of uneven left half (ablation) | PST patch-only · Something-Something V1 | top-1: 45.3 -> 48.6 | Improved |
| use pattern C-9 (9-frame temporal field) instead of A-3 (ablation) | PST patch-only · Something-Something V1 | top-1: 48.6 -> 51.8 | Improved |
| use pattern D-16 instead of C-9 (ablation) | PST patch-only · Something-Something V1 | top-1: 51.8 -> 50.0 | Hurt |
| add TPS to all four stages instead of stage 1 only (ablation) | PST-T · Something-Something V1 | top-1: 47.3 -> 51.8 | Improved |
| add shift back (ablation) | PST-T · Something-Something V1 | top-1: 47.3 -> 51.8 | Improved |
| add alternative shift (ablation) | PST-T · Something-Something V1 | top-1: 46.4 -> 51.8 | Improved |
| add shift of relative position encoding (ablation) | PST-T · Something-Something V1 | top-1: 46.1 -> 51.8 | Improved |
| add channel shift to patch shift (ablation) | PST-T · Something-Something V1 | top-1: 51.8 -> 52.2 | Improved |
| use PST instead of joint spatiotemporal SA | Swin-T · Something-Something V1 | top-1 / memory: 51.5 -> 52.2 top-1; 106G -> 72G FLOPs; 20.2G -> 3.7G memory | Improved |
| use TPS instead of average pooling temporal modeling | DeiT-S · Kinetics-400 | top-1: 73.0 -> 75.3 | Improved |
| use TPS instead of temporal channel shift (TSM) | DeiT-S · Kinetics-400 | top-1: 74.8 -> 75.3 | Improved |
Closest papers by meaning
- Shifted Chunk Transformer for Spatio-Temporal Representational Learning (NeurIPS 2021 · similarity 0.88)
- Video-FocalNets: Spatio-Temporal Focal Modulation for Video Action Recognition (ICCV 2023 · similarity 0.88)
- Spatiotemporal Fusion in 3D CNNs: A Probabilistic View (CVPR 2020 · similarity 0.88)
- Space-time Mixing Attention for Video Transformer (NeurIPS 2021 · similarity 0.88)
- An Efficient Spatio-Temporal Pyramid Transformer for Action Detection (ECCV 2022 · similarity 0.88)
- Shrinking Temporal Attention in Transformers for Video Action Recognition (AAAI 2022 · 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.
