Research map / Video / Video segmentation and tracking
AdaFocus V2: End-to-End Training of Spatial Dynamic Networks for Video Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 AdaFocusV2 with 128² patches instead of VideoIQ | ActivityNet · ActivityNet | mAP: 74.8% -> 79.0% | Helped |
| Use AdaFocusV2 with 128² patches instead of VideoIQ | FCVID · FCVID | mAP: 82.7% -> 85.0% | Helped |
| Use AdaFocusV2 with 128² patches instead of VideoIQ | Mini-Kinetics · Mini-Kinetics | Top-1 Acc.: 72.3% -> 75.4% | Helped |
| Add confidence-based adaptive early exit | AdaFocusV2 · video recognition benchmarks | inference efficiency | Helped |
| Add auxiliary linear-classifier supervision to the global and local encoders (ablation) | AdaFocusV2 end-to-end training · reported experimental settings | recognition performance | Helped |
| Add random-patch diversity augmentation for the local encoder (ablation) | AdaFocusV2 end-to-end training · reported experimental settings | recognition performance | Helped |
| Stop gradients from the policy network before its global-feature inputs (ablation) | AdaFocusV2 end-to-end training · reported experimental settings | recognition performance | Helped |
| Replace random patch selection with the learned patch-selection policy (ablation) | AdaFocusV2-MN2/RN · ActivityNet, after 16 frames | mAP: 73.9% -> 76.4% | Helped |
| Replace center crops with the learned patch-selection policy (ablation) | AdaFocusV2-MN2/RN · ActivityNet, after 16 frames | mAP: 72.6% -> 76.4% | Helped |
| Replace Gaussian patch sampling with the learned patch-selection policy (ablation) | AdaFocusV2-MN2/RN · ActivityNet, after 16 frames | mAP: 73.4% -> 76.4% | Helped |
| Reduce patch size from 176² to 128² (ablation) | AdaFocusV2-TSM · Sth-Sth V1 | GFLOPs: 33.7 -> 18.5 | Helped |
| Reduce patch size from 176² to 128² (ablation) | AdaFocusV2-TSM · Sth-Sth V2 | Top-1 Acc.: 61.3% -> 59.6% | Hurt |
Closest papers by meaning
- AdaFrame: Adaptive Frame Selection for Fast Video Recognition (CVPR 2019 · similarity 0.87)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · similarity 0.87)
- Adapting Shortcut With Normalizing Flow: An Efficient Tuning Framework for Visual Recognition (CVPR 2023 · similarity 0.86)
- 2D or not 2D? Adaptive 3D Convolution Selection for Efficient Video Recognition (CVPR 2021 · similarity 0.85)
- Spatiotemporal Pyramid Network for Video Action Recognition (CVPR 2017 · similarity 0.85)
- STOP: Integrated Spatial-Temporal Dynamic Prompting for Video Understanding (CVPR 2025 · similarity 0.85)
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.