Research map / Video / Action recognition
TransRAC: Encoding Multi-Scale Temporal Correlation With Transformers for Repetitive Action Counting
Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, Shenghua Gao
Check 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 |
|---|---|---|---|
| Use TransRAC with multi-scale temporal correlation and density-map regression | Repetitive action counting · RepCount part-A test | MAE: 0.5267 -> 0.4431 | Improved |
| Use TransRAC rather than RepNet | Repetitive action counting · RepCount part-B | OBO: 0.0025 -> 0.091 | Improved |
| Use TransRAC rather than RepNet | Repetitive action counting · UCFRep | MAE: 0.9985 -> 0.6401 | Improved |
| Replace self-attention correlation with TSM squared Euclidean distance (ablation) | TransRAC · RepCount part-A test | MAE: 0.4431 -> 0.5678 | Hurt |
| Replace the classifier with a density-map regressor (ablation) | ResNet counting model · RepCount part-A test | MAE: 0.9950 -> 0.6905 | Improved |
| Replace the classifier with a density-map regressor (ablation) | SwinT counting model · RepCount part-A test | OBO: 0.118 -> 0.138 | Improved |
| Use only one temporal scale instead of fusing three scales (ablation) | TransRAC · RepCount part-A test | MAE: 0.4431 -> 0.5434 (Scale-4), 0.6595 (Scale-1), 0.6657 (Scale-8) | Hurt |
| Increase the single-scale video sample count from 64 to 128 frames (ablation) | TransRAC single-scale model · RepCount part-A test | MAE: 0.6595 -> 0.6191 | Improved |
| Generate density-map targets with the Gaussian mean at the beginning rather than the middle of a cycle (ablation) | TransRAC density-map predictor · RepCount part-A test | MAE: 0.4936 -> 0.5295 | Hurt |
| Merge density maps centered at beginning, middle, and end rather than use the middle-centered map (ablation) | TransRAC density-map predictor · RepCount part-A test | MAE: 0.4936 -> 0.5142 | Hurt |
| Use the 8-frame scale rather than single-frame scale with 128 sampled frames (ablation) | TransRAC single-scale model · RepCount part-A test | MAE: 0.6191 -> 0.4926 | Improved |
Later papers that compared with this method
4 later accepted papers report TransRAC as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Efficient Action Counting with Dynamic Queries (IJCV 2026 · better on 7 of 7 shared results)
- Action Dubber: Timing Audible Actions via Inflectional Flow (ICML 2025 · better on 5 of 6 shared results)
- Skim then Focus: Integrating Contextual and Fine-grained Views for Repetitive Action Counting (IJCV 2025 · better on 5 of 5 shared results)
- Count What You Want: Exemplar Identification and Few-Shot Counting of Human Actions in the Wild (AAAI 2024 · better on 4 of 4 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
- CountLLM: Towards Generalizable Repetitive Action Counting via Large Language Model (CVPR 2025 · similarity 0.91)
- Efficient Action Counting with Dynamic Queries (IJCV 2026 · similarity 0.91)
- Context-Aware and Scale-Insensitive Temporal Repetition Counting (CVPR 2020 · similarity 0.91)
- Counting Out Time: Class Agnostic Video Repetition Counting in the Wild (CVPR 2020 · similarity 0.89)
- Skim then Focus: Integrating Contextual and Fine-grained Views for Repetitive Action Counting (IJCV 2025 · similarity 0.88)
- Every Moment Counts: Dense Detailed Labeling of Actions in Complex Videos (IJCV 2017 · 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.
