Research map / Architectures and efficiency / CNN design
Global Context Networks
Yue Cao, Jiarui Xu, Stephen Ching-Feng Lin, Fangyun Wei, Han Hu
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 |
|---|---|---|---|
| insert GC block in all residual blocks instead of one (ablation) | GCNet ResNet-50 · COCO val | AP bbox: 38.1 -> 39.4 | Improved |
| add layer norm inside two-layer bottleneck (ablation) | GC block r16 · COCO val | AP bbox: 38.8 -> 39.4 | Improved |
| replace average pooling with attention pooling (ablation) | GC block · COCO val | AP bbox: 39.1 -> 39.4 | Improved |
| use addition instead of scaling fusion (ablation) | GC block · COCO val | AP bbox: 38.3 -> 39.4 | Improved |
| decrease bottleneck ratio from 32 to 4 (ablation) | GC block · COCO val | AP bbox: 39.1 -> 39.9 | Improved |
| insert GC in all stages c3+c4+c5 instead of c3 only (ablation) | GC block · COCO val | AP bbox: 37.9 -> 39.4 | Improved |
| use syncBN instead of fixBN in backbone (ablation) | GCNet · COCO val | AP bbox: 38.5 -> 39.4 | Improved |
| train 24 epochs instead of 12 | GCNet r16 · COCO val | AP bbox: 39.4 -> 39.7 | Improved |
| add GC block to stronger backbone with DCN and Cascade | ResNeXt-101 · COCO val | AP bbox: 47.1 -> 47.9 | Improved |
| add attention pooling and addition fusion versus avg+scale (ablation) | GC block · ImageNet val | top-1 accuracy: 77.14 -> 77.49 | Improved |
| replace NL block with SNL block (ablation) | simplified non-local · COCO val | AP bbox: 38.0 -> 38.1 | No clear effect |
| replace NL with GC block on Kinetics (ablation) | Slow-only R50 · Kinetics val | top-1 accuracy: 75.95 -> 75.85 | No clear effect |
Closest papers by meaning
- LatentGNN: Learning Efficient Non-local Relations for Visual Recognition (ICML 2019 · similarity 0.86)
- Dynamic Context Correspondence Network for Semantic Alignment (ICCV 2019 · similarity 0.84)
- Adaptive Context Network for Scene Parsing (ICCV 2019 · similarity 0.84)
- GLU-Net: Global-Local Universal Network for Dense Flow and Correspondences (CVPR 2020 · similarity 0.84)
- Adaptive Pyramid Context Network for Semantic Segmentation (CVPR 2019 · similarity 0.84)
- GPEN: Global Position Encoding Network for Enhanced Subgraph Representation Learning (ICML 2025 · 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.
