Research map / Video / Action recognition
Few-Shot Action Recognition with Hierarchical Matching and Contrastive Learning
Sipeng Zheng, Shizhe Chen, Qin Jin
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 cross-entropy fine-tuning with spatial-temporal transformer (ablation) | pretrained ResNet-50 · Kinetics 1-shot / SSv2* 1-shot | accuracy: 59.9 -> 62.5; 29.5 -> 37.8 | Improved |
| train with global contrastive loss Lg alone instead of Lce (ablation) | HCL · Kinetics 1-shot | accuracy: 62.5 -> 54.9 | Hurt |
| add global contrastive loss Lg to Lce (ablation) | HCL · Kinetics 1-shot | accuracy: 62.5 -> 66.1 | Improved |
| add temporal contrastive loss with cycle consistency (Lt+Ltc) (ablation) | HCL · SSv2* 1-shot | accuracy: 40.3 -> 45.4 | Improved |
| add spatial contrastive loss with cycle consistency (Ls+Lsc) (ablation) | HCL · Kinetics 1-shot | accuracy: 66.1 -> 72.3 | Improved |
| add both temporal and spatial contrastive losses to Lce+Lg (ablation) | HCL · Kinetics 1-shot / SSv2* 1-shot | accuracy: 66.1 -> 73.7; 40.3 -> 47.3 | Improved |
| add temporal cycle consistency loss Ltc (ablation) | HCL · SSv2* 1-shot / 5-shot | accuracy: -1.5% for 1-shot and -0.9% for 5-shot without Ltc | Improved |
| add spatial cycle consistency loss Lsc (ablation) | HCL · Kinetics 1-shot | accuracy: 72.1 -> 73.1 | Improved |
| add semantic attention in spatial matching (ablation) | HCL · Kinetics 1-shot / 5-shot | accuracy: 71.3 -> 73.7; 84.4 -> 85.8 | Improved |
| add semantic attention in spatial matching (ablation) | HCL · SSv2* 1-shot | accuracy: 46.3 -> 47.3 | No clear effect |
| add temporal and spatial zoom-in matching to global matching (cross-entropy training) (ablation) | Lce-trained model · Kinetics 1-shot | accuracy: 62.5 -> 64.4 | Improved |
| add temporal and spatial zoom-in matching to global matching (HCL training) (ablation) | HCL · Kinetics 1-shot / SSv2* 1-shot | accuracy: 69.1 -> 73.7; 42.9 -> 47.3 | Improved |
Closest papers by meaning
- Matching Compound Prototypes for Few-Shot Action Recognition (IJCV 2024 · similarity 0.90)
- Semi-Supervised Action Recognition With Temporal Contrastive Learning (CVPR 2021 · similarity 0.90)
- Compound Prototype Matching for Few-Shot Action Recognition (ECCV 2022 · similarity 0.90)
- Boosting Few-shot Action Recognition with Graph-guided Hybrid Matching (ICCV 2023 · similarity 0.90)
- A Comprehensive Review of Few-Shot Action Recognition (IJCV 2025 · similarity 0.89)
- Temporal Alignment-Free Video Matching for Few-shot Action Recognition (CVPR 2025 · 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.
