Research map / Video / Action recognition
StartNet: Online Detection of Action Start in Untrimmed Videos
Mingfei Gao, Mingze Xu, Larry S. Davis, Richard Socher, Caiming Xiong
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 LocNet (policy-gradient start localizer) on top of ClsNet (ablation) | StartNet LSTM, TS features · THUMOS'14 offset 1s | p-mAP: 13.9 -> 19.5 | Improved |
| add LocNet (ablation) | StartNet LSTM, RGB features · THUMOS'14 offset 1s | p-mAP: 11.8 -> 15.9 | Improved |
| add LocNet (ablation) | StartNet LSTM, TS features · ActivityNet offset 10s | p-mAP: 13.6 -> 18.0 | Improved |
| add LocNet (ablation) | StartNet LSTM, VGG features · ActivityNet offset 10s | p-mAP: 9.8 -> 13.5 | Improved |
| train LocNet with policy gradient (long-term planning) instead of cross-entropy (ablation) | StartNet LSTM, TS features · THUMOS'14 offset 1s | p-mAP: 17.4 -> 19.5 | Improved |
| policy gradient instead of cross-entropy (ablation) | StartNet LSTM, RGB features · THUMOS'14 offset 1s | p-mAP: 13.7 -> 15.9 | Improved |
| policy gradient instead of cross-entropy (ablation) | StartNet LSTM, TS features · ActivityNet offset 1s | p-mAP: 6.0 -> 8.1 | Improved |
| use LSTM rather than C3D as ClsNet (ablation) | StartNet · THUMOS'14 offset 1s | p-mAP: 6.8 -> 19.5 | Improved |
| use LSTM rather than per-frame CNN as ClsNet (ablation) | StartNet · THUMOS'14 offset 1s | p-mAP: 17.0 -> 19.5 | Improved |
| use two-stream features instead of RGB only | StartNet-PG · THUMOS'14 offset 1s | p-mAP: 15.9 -> 19.5 | Improved |
| two-stage design versus one-stage network predicting starts for all classes (ablation) | StartNet · THUMOS'14 offset 1s, TS features | p-mAP: 10.2% -> 19.5 | Improved |
| feed LocNet action scores rather than low-level image features (ablation) | StartNet · THUMOS'14 offset 1s, TS features | p-mAP: 14.0% -> 19.5 | Improved |
Closest papers by meaning
- Online Detection of Action Start in Untrimmed, Streaming Videos (ECCV 2018 · similarity 0.90)
- One-Shot Action Localization by Learning Sequence Matching Network (CVPR 2018 · similarity 0.89)
- Watch Only Once: An End-to-End Video Action Detection Framework (ICCV 2021 · similarity 0.89)
- Graph Convolutional Module for Temporal Action Localization in Videos (TPAMI 2021 · similarity 0.89)
- Learning To Refactor Action and Co-Occurrence Features for Temporal Action Localization (CVPR 2022 · similarity 0.88)
- Learning to Localize Actions from Moments (ECCV 2020 · similarity 0.88)
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.
