Research map / Video / Action recognition
TempCLR: Temporal Alignment Representation with Contrastive Learning
Yuncong Yang, Jiawei Ma, Shiyuan Huang, Long Chen, Xudong Lin, Guangxing Han, Shih-Fu Chang
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 sequence-level contrastive temporal alignment during zero-shot training | VideoCLIP · CrossTask zero-shot action step localization | recall: 33.9 to 36.9 | Improved |
| use TempCLR pretraining for few-shot recognition | Something-Something V2 · 1-shot action recognition | accuracy: 45.4 to 47.8 | Improved |
| use TempCLR pretraining for few-shot recognition | Something-Something V2 · 5-shot action recognition | accuracy: 64.6 to 67.7 | Improved |
| add temporal-alignment learning | VideoCLIP · DiDeMo clip-caption retrieval | R@1: 16.6 to 17.7 | Improved |
| shuffle clip order within each segment in addition to segment order (ablation) | TempCLR seg-only negatives · CrossTask supervised action step localization | recall: 52.1 to 52.5 | Improved |
| replace seg-unit negatives with within-segment shuffling only (ablation) | TempCLR · CrossTask supervised action step localization | recall: 52.5 to 46.4 | Hurt |
| replace seg-unit negatives with unpaired-video negatives (ablation) | TempCLR · CrossTask supervised action step localization | recall: 52.5 to 48.0 | Hurt |
| replace seg-unit negatives with all-unit shuffling (ablation) | TempCLR · CrossTask supervised action step localization | recall: 52.5 to 49.3 | Hurt |
| add unpaired-video negatives to seg-unit negatives (ablation) | TempCLR · CrossTask supervised action step localization | recall: 52.5 to 52.5 | No clear effect |
| update all parameters instead of norm layers only during zero-shot pretraining (ablation) | TempCLR · CrossTask zero-shot action step localization with OTAM | recall: 36.1 to 36.9 | Improved |
| use DTW instead of OTAM for zero-shot evaluation (ablation) | TempCLR + All Params. · CrossTask zero-shot action step localization | recall: 36.9 to 36.9 | No clear effect |
| use DTW instead of OTAM for full-video retrieval (ablation) | TempCLR · full-video retrieval, background kept | R@1: 72.2 to 70.4 | Hurt |
Later papers that compared with this method
1 later accepted paper reports TempCLR as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Multi-granularity Correspondence Learning from Long-term Noisy Videos (ICLR 2024 · better on 7 of 7 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
- Temporal Alignment Networks for Long-Term Video (CVPR 2022 · similarity 0.89)
- Exploring Temporal Concurrency for Video-Language Representation Learning (ICCV 2023 · similarity 0.89)
- Representation Learning via Global Temporal Alignment and Cycle-Consistency (CVPR 2021 · similarity 0.89)
- TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval (ICLR 2025 · similarity 0.88)
- Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning (CVPR 2024 · similarity 0.87)
- VideoComp: Advancing Fine-Grained Compositional and Temporal Alignment in Video-Text Models (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.
