Research map / Video / Action recognition
Contrast and Order Representations for Video Self-Supervised Learning
Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, Zhiqiang Shen
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 the order task to the contrast-only objective (ablation) | CORPm · Something-Something V2 | top-1 accuracy: 29.7% -> 48.8% | Improved |
| add the contrast task to the order-only objective (ablation) | CORPm · Something-Something V2 | top-1 accuracy: 42.4% -> 48.8% | Improved |
| increase clips per video from 2 to 4 (ablation) | CORPm · Something-Something V2 | top-1 accuracy: 45.4% -> 48.8% | Improved |
| use decoupling attention instead of mixed multi-head attention (ablation) | CORPm · Something-Something V2 | top-1 accuracy: 48.0% -> 48.8% | Improved |
| use 2 perceptron layers instead of 1 (ablation) | CORPm · Something-Something V2 | top-1 accuracy: 48.4% -> 48.8% | Improved |
| add order classification to SimCLR-style contrast (CORPf vs SimCLR) (ablation) | CORPf · Something-Something V2 | top-1 accuracy: 33.9% -> 41.7% | Improved |
| sample more clips per video (CORPm instead of CORPf) (ablation) | CORP · Kinetics-400 (linear eval) | top-1 accuracy: 66.3 % -> 59.1 % | Hurt |
| sample more clips per video (CORPm instead of CORPf) (ablation) | CORP · Something-Something V2 | top-1 accuracy: 41.7% -> 48.8% | Improved |
| increase batch size from 256 to 512 at 200 epochs (ablation) | CORPf · Kinetics-400 (linear eval) | top-1 accuracy: 60.9% -> 64.1% | Improved |
| train 500 instead of 200 epochs at batch size 512 (ablation) | CORPf · Kinetics-400 (linear eval) | top-1 accuracy: 64.1% -> 65.6% | Improved |
| increase batch size from 512 to 1024 (ablation) | CORPf · Kinetics-400 (linear eval) | top-1 accuracy: 66.3 % -> 66.6% | Improved |
| CORPf versus CVRL with 1% labels | CORPf · Kinetics-400 (semi-supervised, 1% labels) | top-1 accuracy: 35.1% -> 34.8% | No clear effect |
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.
- SeCo — SeCo: Exploring Sequence Supervision for Unsupervised Representation Learning (AAAI 2021) · better on 1 of 1 shared result
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
- Self-Supervised Video Representation Learning in a Heuristic Decoupled Perspective (IJCV 2026 · similarity 0.91)
- Frame-Wise Action Representations for Long Videos via Sequence Contrastive Learning (CVPR 2022 · similarity 0.91)
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.90)
- Time-Equivariant Contrastive Video Representation Learning (ICCV 2021 · similarity 0.90)
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.90)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.89)
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.
