Research map / Video / Video segmentation and tracking
Static and Dynamic Concepts for Self-Supervised Video Representation Learning
Rui Qian, Shuangrui Ding, Xian Liu, Dahua Lin
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 |
|---|---|---|---|
| add fidelity and diversity regularizations (ablation) | decoupled concept alignment · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 61.4 -> 68.1 | Improved |
| add local concept contrast L_loc (ablation) | decoupled concept alignment only · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 61.4 -> 67.4 | Improved |
| add local concept contrast L_loc (ablation) | alignment plus L_fid and L_div · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 68.1 -> 72.1 | Improved |
| add local concept contrast L_loc (ablation) | alignment plus L_fid and L_div · HMDB-51 linear probe (K400 pretrain) | top-1 accuracy: 43.2 -> 45.9 | Improved |
| add fidelity and diversity regularizations (ablation) | decoupled concept alignment · HMDB-51 finetune (K400 pretrain) | top-1 accuracy: 44.7 -> 47.9 | Improved |
| add local concept contrast L_loc (ablation) | alignment plus L_fid and L_div · UCF-101 finetune (K400 pretrain) | top-1 accuracy: 80.1 -> 82.1 | Improved |
| use 50+50 concepts instead of 25+25, with L_loc (ablation) | full method · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 70.3 -> 72.1 | Improved |
| increase concepts from 25+25 to 200+200 without L_loc (ablation) | method without L_loc · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 61.2 -> 69.4 | Improved |
| increase concepts from 50+50 to 200+200, with L_loc (ablation) | full method · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 72.1 -> 72.3 | No clear effect |
| add L_loc when concepts are 25+25 (ablation) | method · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 61.2 -> 70.3 | Improved |
| classify with concept latent code q_v instead of global feature v (ablation) | full method · UCF-101 linear probe (K400 pretrain) | top-1 accuracy: 72.1 -> 66.3 | Hurt |
| classify with local feature set F_v instead of global feature v (ablation) | full method · UCF-101 / Diving-48 linear | top-1 accuracy: UCF-101 72.1 -> 72.7; Diving-48 73.4 -> 72.5 | Mixed |
Closest papers by meaning
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.89)
- Suppressing Static Visual Cues via Normalizing Flows for Self-Supervised Video Representation Learning (AAAI 2022 · similarity 0.89)
- Contrast and Order Representations for Video Self-Supervised Learning (ICCV 2021 · similarity 0.88)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.87)
- PooDLe🐩: Pooled and dense self-supervised learning from naturalistic videos (ICLR 2025 · similarity 0.87)
- When the Future Becomes the Past: Taming Temporal Correspondence for Self-supervised Video Representation Learning (CVPR 2025 · 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.
