Research map / Video / Action recognition
Self-Supervised Spatiotemporal Representation Learning by Exploiting Video Continuity
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 |
|---|---|---|---|
| Use all three continuity-aware pretext tasks jointly (ablation) | CPNet · UCF101, HMDB51, Diving48 | top-1 accuracy: 80.7 vs. 78.4; 51.8 vs. 47.2; 72.5 vs. 71.0 | Helped |
| Add discontinuity localization (ablation) | CPNet · UCF101 | top-1 accuracy: 74.3 -> 78.4 | Helped |
| Add continuity justification (ablation) | CPNet · UCF101 | top-1 accuracy: 75.5 -> 78.6 | Helped |
| Add missing section approximation (ablation) | CPNet · UCF101 | top-1 accuracy: 78.4 -> 78.6 | Helped |
| Add triplet loss to missing-section contrastive learning (ablation) | CPNet · UCF101 | top-1 accuracy: 80.0 -> 80.7 | Helped |
| Set missing-section length to 8 rather than 4 or 16 (ablation) | CPNet · UCF101, HMDB51, Diving48 | top-1 accuracy: 80.1, 49.9, 72.1 -> 80.7, 51.8, 72.5; length 16: 79.8, 47.3, 71.6 | Helped |
| Add discontinuity localization to prior self-supervised methods | PacePred, VCOP, and MOCO · UCF101 and HMDB51 | top-1 accuracy: PacePred: +0.6 to +3.9 UCF101 and +1.7 to +3.3 HMDB51; VCOP: +1.3 to +2.2 UCF101 and +0.5 to +3.2 HMDB51; MOCO: +0.6 to | Helped |
| Pretrain CPNet on K400 rather than UCF101 | CPNet with R(2+1)D · UCF101 and HMDB51 | top-1 accuracy: 81.8 -> 83.8; 51.2 -> 57.1 | Helped |
| Evaluate K400-pretrained CPNet at 15 fps rather than 25 fps (ablation) | CPNet with R(2+1)D · UCF101 and HMDB51 | top-1 accuracy: 82.1 -> 83.8; 53.2 -> 57.1 | Helped |
| Use feature-space contrastive missing-section approximation rather than RGB reconstruction (ablation) | CPNet versus CRNet · UCF101/HMDB51 video retrieval, R(2+1)D R@1 | R@1: 35.3/14.0 vs. 22.9/12.0 | Helped |
Closest papers by meaning
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.88)
- A Large-Scale Study on Unsupervised Spatiotemporal Representation Learning (CVPR 2021 · similarity 0.88)
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.88)
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.87)
- Video Playback Rate Perception for Self-Supervised Spatio-Temporal Representation Learning (CVPR 2020 · 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.