Research map / Video / Action recognition
Cycle-Contrast for Self-Supervised Video Representation Learning
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 MSE video-frame loss with cycle-contrastive loss (L_v) (ablation) | CCL pretraining · UCF101 fine-tune | top-1 accuracy: 52.8 -> 63.3 | Helped |
| add frame-level forward contrast L_f to L_v (ablation) | CCL pretraining · UCF101 fine-tune | top-1 accuracy: 63.3 -> 67.6 | Helped |
| add penalization term P (ablation) | CCL pretraining · UCF101 fine-tune | top-1 accuracy: 67.6 -> 69.4 | Helped |
| add penalization term P (ablation) | CCL pretraining · HMDB51 fine-tune | top-1 accuracy: 35.2 -> 37.8 | Helped |
| add frame-level forward contrast L_f to L_v (ablation) | CCL pretraining · HMDB51 fine-tune | top-1 accuracy: 31.1 -> 35.2 | Helped |
| cycle-contrastive loss instead of MSE for video-frame retrieval (ablation) | CCL · UCF101 F=>V retrieval | top-1 accuracy: 15.3 -> 26.1 | Helped |
| fine-tune all layers end-to-end instead of FC only | CCL pretrained R3D · MMAct cross-session | f-measure: 59.1 -> 65.1 | Helped |
| CCL pretraining vs random init, only FC fine-tuned | R3D-18+1 · MMAct cross-session | f-measure: 62.2 -> 59.1 | Hurt |
| CCL pretraining on Kinetics vs random init, fine-tuned end-to-end | R3D-18+1 · UCF101 | top-1 accuracy: 44.7 -> 69.4 | Helped |
Closest papers by meaning
- Cross-Architecture Self-Supervised Video Representation Learning (CVPR 2022 · similarity 0.92)
- Frame-Wise Action Representations for Long Videos via Sequence Contrastive Learning (CVPR 2022 · similarity 0.89)
- Learning Correspondence From the Cycle-Consistency of Time (CVPR 2019 · similarity 0.88)
- Self-supervised Co-Training for Video Representation Learning (NeurIPS 2020 · similarity 0.88)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · similarity 0.87)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.86)
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.