Research map / Architectures and efficiency / CNN design
Local Relation Networks for Image Recognition
Han Hu, Zheng Zhang, Zhenda Xie, Stephen Lin
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 |
|---|---|---|---|
| restrict aggregation to a local 7x7 neighborhood instead of full image (ablation) | LR-Net-26 (no geometric prior) · ImageNet val | top-1 accuracy: 50.7 -> 71.9 | Improved |
| restrict aggregation to a local 7x7 neighborhood instead of full image (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 68.4 -> 75.7 | Improved |
| add geometric prior term (small network on relative position) (ablation) | LR-Net-26 7x7 · ImageNet val | top-1 accuracy: 71.9 -> 75.7 | Improved |
| compute geometric prior with small network instead of directly learned values (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 72.3 -> 75.4 at 9x9 | Improved |
| increase kernel size from 3x3 to 7x7 (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 73.6 -> 74.9 -> 75.7 | Improved |
| increase convolution kernel size from 5x5 to 9x9 (ablation) | ResNet-26 · ImageNet val | top-1 accuracy: 73.0 -> 71.4 | Hurt |
| use scalar query/key instead of 16-dim vectors (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 73.7 -> 75.7 | Improved |
| share channels (m=8) in aggregation weights (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 75.3 -> 75.7 | No clear effect |
| add softmax normalization over spatial scope (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 74.8 -> 75.7 | Improved |
| change appearance composability function (squared diff, abs diff, product) (ablation) | LR-Net-26 · ImageNet val | top-1 accuracy: 75.7, 75.5, 75.7 | No clear effect |
| replace 3x3 convolution with non-local module | ResNet-26 · ImageNet val | top-1 accuracy: 72.8 -> 47.7 | Hurt |
| add non-local modules after every block | ResNet-26 · ImageNet val | top-1 accuracy: 72.8 -> 73.4 | Improved |
Closest papers by meaning
- PARN: Position-Aware Relation Networks for Few-Shot Learning (ICCV 2019 · similarity 0.87)
- LatentGNN: Learning Efficient Non-local Relations for Visual Recognition (ICML 2019 · similarity 0.87)
- Broadcasting Convolutional Network for Visual Relational Reasoning (ECCV 2018 · similarity 0.86)
- MR-NET: Exploiting Mutual Relation for Visual Relationship Detection (AAAI 2019 · similarity 0.86)
- Relationship Proposal Networks (CVPR 2017 · similarity 0.86)
- Relation Networks for Object Detection (CVPR 2018 · similarity 0.86)
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.
