Research map / Video / Action recognition
Context-Aware RCNN: A Baseline for Action Detection in Videos
Jianchao Wu, Zhanghui Kuang, Limin Wang, Wayne Zhang, Gangshan Wu
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 |
|---|---|---|---|
| replace RoI Pooling with crop-and-resize actor features (32x2) (ablation) | actor-feature action detector · AVA val, actor features only | mAP: 22.1 -> 25.0 | Improved |
| replace RoI Pooling with crop-and-resize for extra small boxes (ablation) | actor-feature action detector · AVA val, extra small boxes | mAP: 12.9 -> 16.9 | Improved |
| replace RoI Pooling with crop-and-resize for extra large boxes (ablation) | actor-feature action detector · AVA val, extra large boxes | mAP: 22.2 -> 22.0 | No clear effect |
| replace RoI Pooling with crop-and-resize when one actor in frame (ablation) | actor-feature action detector · AVA val, 1 actor per frame | mAP: 29.2 -> 35.2 | Improved |
| replace RoI Pooling with crop-and-resize when 9 to 37 actors in frame (ablation) | actor-feature action detector · AVA val, 9-37 actors per frame | mAP: 17.2 -> 22.9 | Improved |
| increase actor input resolution from 112x112 to 224x224 (ablation) | Context-Aware RCNN · AVA val, actor features only | mAP: 21.4 -> 25.0 | Improved |
| increase frames from 8x8 to 32x2 sampling with crop-and-resize (ablation) | Crop+Resize · AVA val, actor features only | mAP: 23.1 -> 25.0 | Improved |
| expand actor box by max scale 1.5 versus 1.2 (ablation) | Crop+Resize · AVA val, actor features only | mAP: 24.1 -> 25.0 | Improved |
| expand actor box to max scale 2.5 versus 1.5 (ablation) | Crop+Resize · AVA val, actor features only | mAP: 25.0 -> 24.3 | Hurt |
| expand actor box with scale 1.5 in RoI Pooling pipeline (ablation) | RoI Pooling · AVA val, actor features only | mAP: 22.1 -> 21.6 | Hurt |
| add scene feature by concatenation (ablation) | Context-Aware RCNN · AVA val, 16x4 | mAP: 24.7 -> 25.7 | Improved |
| use transformer attention to scene features instead of concat (ablation) | Context-Aware RCNN · AVA val, 16x4 | mAP: 25.7 -> 25.8 | No clear effect |
Closest papers by meaning
- A Structured Model for Action Detection (CVPR 2019 · similarity 0.89)
- Watch Only Once: An End-to-End Video Action Detection Framework (ICCV 2021 · similarity 0.88)
- G-TAD: Sub-Graph Localization for Temporal Action Detection (CVPR 2020 · similarity 0.88)
- Efficient Video Action Detection with Token Dropout and Context Refinement (ICCV 2023 · similarity 0.88)
- VideoNet: A Large-Scale Dataset for Domain-Specific Action Recognition (CVPR 2026 · similarity 0.88)
- CycleACR: Cycle Modeling of Actor-Context Relations for Video Action Detection (TPAMI 2025 · 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.
