Research map / Video / Action recognition
Self-supervised Motion Representation via Scattering Local Motion Cues
Yuan Tian, Zhaohui Che, Wenbo Bao, Guangtao Zhai, Zhiyong Gao
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 nearest upsampling with learnable CGMUL (ablation) | motion map upsampling · UCF-Flow | l1 error (lower better): 0.042 -> 0.024 | Improved |
| replace bilinear upsampling with learnable CGMUL (ablation) | motion map upsampling · UCF-Flow | l1 error (lower better): 0.037 -> 0.024 | Improved |
| add CGMUL-upsampled HR motion map (vs bilinear) on top of LR (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 88.2% -> 90.6% | Improved |
| add nearest or bilinear upsampled HR motion map on top of LR (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 88.2% -> 88.2% | No clear effect |
| add LR motion map to the backbone (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 84.4% -> 88.2% | Improved |
| train backbone with motion representation learning (no extra inference cost) (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 84.4 -> 86.5 | Improved |
| add HR motion pipeline fused with backbone (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 88.2 -> 89.4 | Improved |
| add LR+HR fusion instead of HR alone (ablation) | 3D-ResNet18 · UCF101 | top-1 accuracy: 89.4 -> 90.6 | Improved |
| use shallower 3D-ResNet18 with LR+HR motion rather than 3D-ResNet34 baseline | action recognition · UCF101 | top-1 accuracy: 87.7 -> 90.6 | Improved |
| increase motion-map upsampling scale factor to 4 from 1 (ablation) | CGM-Net · action recognition (Fig. 6) | top-1 accuracy: 6.2% | Improved |
| increase upsampling scale factor beyond 8 (e.g. 16) (ablation) | CGM-Net · motion estimation / action recognition (Fig. 6) | l1 error / top-1 | Hurt |
| add CGM-Net multi-task motion learning on 3D-ResNeXt101 vs DynamoNet | 3D-ResNeXt101 · Kinetics-400 | top-1 accuracy: 68.2 -> 76.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- DynamoNet (64) — DynamoNet: Dynamic Action and Motion Network (ICCV 2019) · better on 0 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
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.90)
- Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion (AAAI 2021 · similarity 0.89)
- Self-Supervised Representation Learning From Flow Equivariance (ICCV 2021 · similarity 0.87)
- Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition (CVPR 2018 · similarity 0.87)
- FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025 · similarity 0.86)
- Self-Supervised Learning of Motion Concepts by Optimizing Counterfactuals (NeurIPS 2025 · similarity 0.85)
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.
