Research map / Video / Action recognition
AdaFocusV3: On Unified Spatial-Temporal Dynamic Video Recognition
Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, Gao Huang
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 learned temporal policy (ablation) | AdaFocusV3 · ActivityNet (1 cube) | mAP: 45.5% -> 51.9% | Improved |
| add learned spatial policy (ablation) | AdaFocusV3 · ActivityNet (1 cube) | mAP: 50.1% -> 51.9% | Improved |
| use deep-feature gradient estimation instead of pixel-based (ablation) | AdaFocusV3 · ActivityNet (1 cube) | mAP: 50.6% -> 51.9% | Improved |
| use deep-feature gradient estimation instead of pixel-based (ablation) | AdaFocusV3 · ActivityNet (8 cubes) | mAP: 75.0% -> 75.4% | Improved |
| add learned temporal policy (ablation) | AdaFocusV3 · ActivityNet (8 cubes) | mAP: 74.6% -> 75.4% | Improved |
| add learned spatial policy (ablation) | AdaFocusV3 · ActivityNet (8 cubes) | mAP: 74.6% -> 75.4% | Improved |
| unified spatial-temporal modeling instead of separate modules (V3 vs V2+) | AdaFocusV3 · ActivityNet at 79.0% mAP | GFLOPs/video: 24.4 -> 16.8 | Improved |
| unified spatial-temporal modeling instead of separate modules (V3 vs V2+) | AdaFocusV3 · FCVID at 85.0% mAP | GFLOPs/video: 24.4 -> 14.4 | Improved |
| shorten cube temporal length from 12 to 8 (ablation) | AdaFocusV3 (TSM) · Sth-Sth V1 | top-1 accuracy: 47.0% -> 46.1% | Hurt |
| shorten cube temporal length from 12 to 8 (ablation) | AdaFocusV3 (TSM) · Diving48 | top-1 accuracy: 80.4% -> 79.3% | Hurt |
| enlarge cube temporal length from 12 to 16 (ablation) | AdaFocusV3 (TSM) · Sth-Sth V1 | GFLOPs: 14.0 -> 15.9 | Hurt |
| apply AdaFocusV3 instead of TSM+ baseline | TSM+ · Diving48 | GFLOPs/video: 35.1 -> 6.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OCSampler — OCSampler: Compressing Videos to One Clip With Single-Step Sampling (CVPR 2022) · 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
- Uni-AdaFocus: Spatial-Temporal Dynamic Computation for Video Recognition (TPAMI 2024 · similarity 0.95)
- AdaFocus V2: End-to-End Training of Spatial Dynamic Networks for Video Recognition (CVPR 2022 · similarity 0.94)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · similarity 0.89)
- Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification (ECCV 2018 · similarity 0.89)
- Glance and Focus Networks for Dynamic Visual Recognition (TPAMI 2022 · similarity 0.88)
- A Dynamic Frame Selection Framework for Fast Video Recognition (TPAMI 2020 · 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.
