Research map / Video / Video segmentation and tracking
Compressed Video Contrastive Learning
Yuqi Huo, Mingyu Ding, Haoyu Lu, Nanyi Fei, Zhiwu Lu, Ji-Rong Wen, Ping Luo
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 |
|---|---|---|---|
| downsample optical flow to half resolution (ablation) | optical flow init · UCF101 | linear probe top-1: 51.3 -> 66.8 | Hurt |
| add cross guidance pre-training with RGB to motion vector view | MV encoder · UCF101 | linear probe top-1: 65.0 -> 73.8 | Improved |
| add cross guidance pre-training with RGB to optical flow view | OF encoder · UCF101 | linear probe top-1: 66.8 -> 73.1 | Improved |
| use CoViAR-style P-frame motion vectors instead of on-the-fly decoded ones (ablation) | initialization · UCF101 | linear probe top-1: 63.4 -> 65.0 | No clear effect |
| replace CoCLR with MVCGC cross guidance (RGB+MV) | RGB network · UCF101 | linear probe top-1: 67.4 -> 73.1 | Improved |
| replace CoCLR with MVCGC cross guidance (RGB+MV) | RGB network · UCF101 | retrieval top-1: 53.8 -> 60.8 | Improved |
| replace CoCLR with MVCGC using optical flow as second view | RGB network · UCF101 | linear probe top-1: 70.2 -> 70.6 | Improved |
| remove initialization stage before cross guidance (ablation) | MVCGC · UCF101 | linear probe top-1: 73.1 -> 65.0 | Hurt |
| use union instead of intersection of top-k candidates for positives (ablation) | MVCGC · UCF101 | linear probe top-1: 73.1 -> 71.3 | Hurt |
| reduce top-k from 5 to 1 (ablation) | MVCGC · UCF101 | linear probe top-1: 73.1 -> 71.5 | Hurt |
| increase top-k from 5 to 10 (ablation) | MVCGC · UCF101 | linear probe top-1: 73.1 -> 72.7 | Mixed |
| increase top-k from 5 to 20 (ablation) | MVCGC · UCF101 | linear probe top-1: 73.1 -> 71.5 | Hurt |
Closest papers by meaning
- Motion Sensitive Contrastive Learning for Self-Supervised Video Representation (ECCV 2022 · similarity 0.89)
- Motion-Focused Contrastive Learning of Video Representations (ICCV 2021 · similarity 0.88)
- Contrast and Order Representations for Video Self-Supervised Learning (ICCV 2021 · similarity 0.88)
- Cross-Architecture Self-Supervised Video Representation Learning (CVPR 2022 · similarity 0.87)
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.87)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · 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.
