Research map / Video / Action recognition
A Closer Look at Spatiotemporal Convolutions for Action Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Replace full 3D convolutions with factorized spatial-then-temporal (2+1)D blocks | 18-layer ResNet · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 68.0 | Helped |
| Use temporal convolutions instead of framewise 2D processing | ResNet · Kinetics validation, 16-frame clips | Video@1: 60.5 -> 64.2 | Helped |
| Use temporal convolutions instead of collapsing frames into channels | ResNet · Kinetics validation, 16-frame clips | Video@1: 58.9 -> 64.2 | Helped |
| Use 2D convolutions in the deeper groups, retaining 3D convolutions in earlier groups (ablation) | R3D · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 65.1 | Helped |
| Move 2D processing to earlier groups, using 3D convolutions only in later groups (ablation) | R3D · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 65.1 | Helped |
| Use 2D convolutions in group 5 (ablation) | R3D · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 65.1 | Helped |
| Use 2D convolutions from group 2 onward (ablation) | R3D · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 64.2 | No clear effect |
| Use 2D convolutions in groups 1–4 and 3D convolutions only in group 5 (ablation) | R3D · Kinetics validation, 16-frame clips | Video@1: 64.2 -> 63.1 | Hurt |
| Train R(2+1)D on longer clips rather than 8-frame clips (ablation) | 18-layer R(2+1)D · Kinetics validation | Video@1: 64.8 -> 69.4 | Helped |
| Test an 8-frame-trained model on 32-frame clips instead of its training clip length (ablation) | 18-layer R(2+1)D · Kinetics validation | Video@1: 64.8 -> 59.0 | Hurt |
| Fine-tune the 8-frame model on 32-frame clips (ablation) | 18-layer R(2+1)D · Kinetics validation | Video@1: 64.8 -> 68.0 | Helped |
| Pretrain on Sports-1M rather than train from scratch (ablation) | R(2+1)D-RGB · Kinetics | top1: 72.0 -> 74.3 | Helped |
Closest papers by meaning
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.93)
- Spatiotemporal Residual Networks for Video Action Recognition (NeurIPS 2016 · similarity 0.92)
- Spatiotemporal Fusion in 3D CNNs: A Probabilistic View (CVPR 2020 · similarity 0.91)
- Collaborative Spatiotemporal Feature Learning for Video Action Recognition (CVPR 2019 · similarity 0.91)
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.90)
- Spatiotemporal Pyramid Network for Video Action Recognition (CVPR 2017 · similarity 0.90)
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.