Research map / Video / Action recognition
Motion-Aware Contrastive Video Representation Learning via Foreground-Background Merging
Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, Hongkai Xiong
Check 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 FAME by composing a clip's moving foreground with another video's background | MoCo video contrastive pretraining · UCF101 and HMDB51 finetune, Kinetics-400 pretraining | Top-1 accuracy: 82.0 -> 84.8; 51.2 -> 53.5 | Improved |
| Add FAME foreground-background merging | I3D pretrained on Kinetics-400 · Diving48 V2 | Top-1 accuracy: 62.4 -> 72.9 | Improved |
| Reduce the foreground area ratio from the beta=1.0 baseline to beta=0.3 or 0.5 (ablation) | FAME with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 81.2; 45.5 -> 52.6 | Improved |
| Apply FAME to both contrastive branches rather than only one (ablation) | FAME with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 81.2 vs 81.2; 51.4 vs 52.6 | No clear effect |
| Use an intra-video background rather than no background replacement (ablation) | FAME with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 77.4(1.6$\uparrow$); 45.5 -> 47.6(2.1$\uparrow$) | Improved |
| Replace intra-video background with a different video's background (ablation) | FAME with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 81.2(5.4$\uparrow$); 45.5 -> 52.6(7.1$\uparrow$) | Improved |
| Replace the FAME foreground mask with a Gaussian mask (ablation) | MoCo with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 77.9; 45.5 -> 46.4 | Improved |
| Replace the FAME foreground mask with the frame-difference seed region (ablation) | MoCo with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 80.4; 45.5 -> 51.3 | Improved |
| Replace the FAME foreground mask with a coarse 4 x 4 motion grid (ablation) | MoCo with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 81.5; 45.5 -> 51.5 | Mixed |
| Use the refined FAME mask instead of the grid mask (ablation) | MoCo with I3D pretrained on Kinetics-400 · UCF101 and HMDB51 finetune | Top-1 accuracy: 86.5 -> 88.6; 58.7 -> 61.1 | Improved |
| Replace the no-FAME baseline with FAME using the color-statistics mask (ablation) | MoCo with I3D · UCF101 and HMDB51 finetune | Top-1 accuracy: 75.8 -> 81.2; 45.5 -> 52.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- RSPNet — RSPNet: Relative Speed Perception for Unsupervised Video Representation Learning (AAAI 2021) · better on 1 of 2 shared results
- ASCNet — ASCNet: Self-Supervised Video Representation Learning With Appearance-Speed Consistency (ICCV 2021) · better on 2 of 2 shared results
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 0.86)
- Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion (AAAI 2021 · similarity 0.86)
- Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground (ICCV 2023 · similarity 0.86)
- Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision (NeurIPS 2025 · similarity 0.85)
- DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025 · similarity 0.85)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.84)
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.
