Research map / Video / Action recognition
A Coarse-to-Fine Framework for Resource Efficient Video Recognition
Zuxuan Wu, Hengduo Li, Yingbin Zheng, Caiming Xiong, Yu–Gang Jiang, Larry Steven Davis
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 synchronization of fine LSTM with coarse LSTM states (ablation) | LiteEval · FCVID | mAP: 65.7% -> 80.0% | Improved |
| use Gumbel-Softmax instead of REINFORCE for gating (ablation) | LiteEval · FCVID | mAP and GFLOPs: 74.2% -> 80.0%; 245.9 -> 94.3 GFLOPs | Improved |
| use Gumbel-Softmax instead of REINFORCE for gating (ablation) | LiteEval · ActivityNet | mAP and GFLOPs: 65.2% -> 72.7%; 269.3 -> 95.1 GFLOPs | Improved |
| replace uniform sampling of 25 frames with coarse-to-fine gating | LiteEval · FCVID | GFLOPs: 195.5 -> 94.3 GFLOPs at 80.0% mAP | Improved |
| replace uniform sampling of 25 frames with coarse-to-fine gating | LiteEval · ActivityNet | mAP: 70.0% -> 72.7%; 195.5 -> 95.1 GFLOPs | Improved |
| lower fine-feature usage target gamma from 0.05 to 0.01 (ablation) | LiteEval · FCVID | mAP: 80.0% -> 78.8%; 94.3 -> 75.4 GFLOPs | Hurt |
| raise gamma from 0.05 to 0.10 (ablation) | LiteEval · FCVID | mAP: 80.0% -> 80.1%; 94.3 -> 139.0 GFLOPs | No clear effect |
| raise gamma from 0.01 to 0.03 (ablation) | LiteEval · FCVID | mAP: 78.8% -> 79.7%; 75.4 -> 82.1 GFLOPs | Improved |
| increase coarse LSTM hidden units from 64 to 512 (ablation) | LiteEval · FCVID | mAP: 76.9% -> 80.0% | Improved |
| increase coarse LSTM hidden units from 256 to 512 (ablation) | LiteEval · FCVID | mAP: 78.3% -> 80.0% | Improved |
| compare with FrameGlimpse and FastForward agents that jump to future frames | LiteEval · FCVID | mAP: 71.2% -> 80.0%; 67.6% -> 80.0% | Improved |
Closest papers by meaning
- LiteEval: A Coarse-to-Fine Framework for Resource Efficient Video Recognition (NeurIPS 2019 · similarity 0.99)
- AdaFrame: Adaptive Frame Selection for Fast Video Recognition (CVPR 2019 · similarity 0.85)
- Principles of Visual Tokens for Efficient Video Understanding (ICCV 2025 · similarity 0.85)
- Uni-AdaFocus: Spatial-Temporal Dynamic Computation for Video Recognition (TPAMI 2024 · similarity 0.85)
- SlowFast Networks for Video Recognition (ICCV 2019 · similarity 0.84)
- LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute (NeurIPS 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.
