Research map / Video / Action recognition
Composable Augmentation Encoding for Video Representation Learning
Chen Sun, Arsha Nagrani, Yonglong Tian, Cordelia Schmid
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 time shift encoding to the transformer head (ablation) | SimCLR++ · SSv1 linear eval | top-1 accuracy: 26.5 -> 31.2 | Improved |
| add temporal augmentation (frames at different times) to MLP-head SimCLR (ablation) | SimCLR · SSv1 linear eval | top-1 accuracy: 17.1 -> 26.4 | Improved |
| add MLP head on top of the transformer head without encodings (ablation) | SimCLR++ · SSv1 linear eval | top-1 accuracy: 26.5 -> 24.6 | Hurt |
| add crop encoding with dropout (ablation) | SimCLR++ · SSv1 linear eval | top-1 accuracy: 26.5 -> 28.1 | Improved |
| add dropout on crop encoding (ablation) | CATE crop encoding · SSv1 linear eval | top-1 accuracy: 27.2 -> 28.1 | Improved |
| encode time shift delta_t instead of only arrow of time (ablation) | CATE time encoding · SSv1 linear eval | top-1 accuracy: 28.1 -> 31.3 | Improved |
| add dropout on time encoding (ablation) | CATE time encoding · SSv1 linear eval | top-1 accuracy: 31.3 -> 31.2 | No clear effect |
| compose crop encoding with time encoding (ablation) | CATE · SSv1 linear eval | top-1 accuracy: 31.2 -> 32.2 | Improved |
| compose crop encoding with time encoding (ablation) | CATE · SSv2 linear eval | top-1 accuracy: 42.3 -> 43.5 | Improved |
| use 4 transformer layers instead of 1 (ablation) | CATE · SSv1 linear eval | top-1 accuracy: 26.9 -> 31.2 | Improved |
| use 8 transformer layers instead of 4 (ablation) | CATE · SSv1 linear eval | top-1 accuracy: 31.2 -> 31.3 | No clear effect |
| pretrain 800 epochs instead of 500 (ablation) | CATE · SSv1, UCF101, HMDB51 | accuracy: 32.2 -> 33.1; 84.3 -> 83.6; 53.6 -> 53.0 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TaCo — TACo: Token-Aware Cascade Contrastive Learning for Video-Text Alignment (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
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.90)
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.89)
- Time-Equivariant Contrastive Video Representation Learning (ICCV 2021 · similarity 0.88)
- Probabilistic Representations for Video Contrastive Learning (CVPR 2022 · similarity 0.88)
- Motion-Focused Contrastive Learning of Video Representations (ICCV 2021 · similarity 0.88)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 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.
