Research map / Video / Action recognition
Grouped Spatial-Temporal Aggregation for Efficient Action Recognition
Chenxu Luo, Alan L. Yuille
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 |
|---|---|---|---|
| replace one group of 3D group convolution with spatial-only convolution (GST-Large vs C3D groups=2) | C3D groups=2 · Something-Something v1 val | top-1 accuracy: 45.1 -> 47.7 | Improved |
| parallel spatial/temporal channel split instead of cascaded P3D decoupling | P3D · Something-Something v1 val | top-1 accuracy: 45.7 -> 47.7 | Improved |
| same parallel split on v2 | P3D · Something-Something v2 val | top-1 accuracy: 59.8 -> 62.0 | Improved |
| GST-Large versus full 3x3x3 C3D with fewer parameters | C3D 3x3x3 · Something-Something v1 val | top-1 accuracy: 46.2 -> 47.7 | Improved |
| add GST temporal channels to 2D ResNet-50 at equal parameters | C2D · Something-Something v1 val | top-1 accuracy: 20.4 -> 46.7 | Improved |
| reduce temporal channel ratio alpha from 1/2 to 1/4 (ablation) | GST · Something-Something v1 val | top-1 accuracy: 46.7 -> 47.0 | Improved |
| reduce temporal channel ratio alpha from 1/4 to 1/8 (ablation) | GST · Something-Something v2 val | top-1 accuracy: 61.6 -> 60.7 | Hurt |
| reduce temporal channel ratio alpha from 1/2 to 1/8 (ablation) | GST · Something-Something v1 val | top-1 accuracy: 46.7 -> 46.7 | No clear effect |
| increase input frames from 8 to 16 | GST · Something-Something v1 val | top-1 accuracy: 47.0 -> 48.6 | Improved |
| increase input frames from 8 to 16 | GST · Something-Something v2 val | top-1 accuracy: 61.6 -> 62.6 | Improved |
| use two clips at test time instead of one | GST · Something-Something v1 val | top-1 accuracy: 47.0 -> 47.6 | Improved |
| GST instead of P3D backbone | ResNet-50 · Diving48 | accuracy: 32.4 -> 38.8 | Improved |
Closest papers by meaning
- Collaborative Spatiotemporal Feature Learning for Video Action Recognition (CVPR 2019 · similarity 0.88)
- Spatiotemporal Joint Filter Decomposition in 3D Convolutional Neural Networks (NeurIPS 2021 · similarity 0.88)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · similarity 0.87)
- StNet: Local and Global Spatial-Temporal Modeling for Action Recognition (AAAI 2019 · similarity 0.87)
- TEA: Temporal Excitation and Aggregation for Action Recognition (CVPR 2020 · similarity 0.87)
- Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification (ECCV 2018 · 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.
