Research map / Video / Action recognition
Extreme Low-Resolution Action Recognition with Confident Spatial-Temporal Attention Transfer
Yucai Bai, Qin Zou, Xieyuanli Chen, Lingxi Li, Zhengming Ding, Long Chen
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 spatial attention transfer (SAT) to the eLR student (ablation) | Resnet-34 student · UCF101 (12x16) | Prec@1 (%): 42.79% -> 44.42% | Improved |
| add temporal attention transfer (TAT) to the eLR student (ablation) | Resnet-34 student · UCF101 (12x16) | Prec@1 (%): 42.79% -> 45.48% | Improved |
| combine SAT and TAT (STAT) (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 44.42% -> 47.32% | Improved |
| add temporal attention transfer on top of spatial transfer (ablation) | R2Plus1D-18 · UCF101 (12x16) | Prec@1 (%): 52.03% -> 52.72% (TAT alone 50.90%) | Mixed |
| add confidence score to STAT (CSTAT) (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 47.32% -> 49.03% | Improved |
| add rolling weights to CSTAT (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 49.03% -> 49.97% | Improved |
| add unsupervised training to CSTAT+rc (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 49.97% -> 50.78% | Improved |
| add unsupervised training to CSTAT+rc (ablation) | R2Plus1D-18 · UCF101 (12x16) | Prec@1 (%): 54.78% -> 55.31% | Improved |
| use MLP-based confidence map instead of difference-based (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 47.71% -> 49.53% | Improved |
| use gradient-based confidence map over baseline (ablation) | R2Plus1D-18 · UCF101 (12x16) | Prec@1 (%): 52.72% -> 53.51% | Improved |
| use rolling weights instead of best fixed weights (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 47.01% -> 47.18% | Improved |
| use increasing SAT layer weights 1/15-2/15-4/15-8/15 instead of equal (ablation) | Resnet-34 · UCF101 (12x16) | Prec@1 (%): 46.12% -> 45.43% (at external 4/7-1/7-2/7 versus equal weights) | Hurt |
Closest papers by meaning
- Privacy-Preserving Human Activity Recognition from Extreme Low Resolution (AAAI 2017 · similarity 0.88)
- Extreme Low Resolution Activity Recognition With Multi-Siamese Embedding Learning (AAAI 2018 · similarity 0.87)
- Memory-Augmented Non-Local Attention for Video Super-Resolution (CVPR 2022 · similarity 0.84)
- HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.84)
- Fast Temporal Activity Proposals for Efficient Detection of Human Actions in Untrimmed Videos (CVPR 2016 · similarity 0.84)
- DarkShake-DVS: Event-based Human Action Recognition under Low-light and Shaking Camera Conditions (CVPR 2026 · similarity 0.83)
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.
