Research map / Architectures and efficiency / Attention and transformers
Contextual Transformer Networks for Visual Recognition
Yehao Li, Ting Yao, Yingwei Pan, Tao Mei
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 dynamic self-attention context to static context (ablation) | CoTNet-50 · ImageNet-1k val (default training, CoTNet-50) | top-1 accuracy: 77.1 -> 78.5 | Improved |
| fuse static and dynamic contexts by linear summation (ablation) | CoTNet-50 · ImageNet-1k val (default training, CoTNet-50) | top-1 accuracy: 78.5 -> 78.7 | Improved |
| fuse static and dynamic contexts via attention instead of summation (ablation) | CoTNet-50 · ImageNet-1k val (default training, CoTNet-50) | top-1 accuracy: 78.7 -> 79.2 | Improved |
| replace 3x3 convs in res4 as well as res5 (ablation) | CoTNet-50 · ImageNet-1k val (default training, CoTNet-50) | top-1 accuracy: 78.5 -> 79.0 | Improved |
| replace 3x3 convs in res2 and res3 as well (ablation) | CoTNet-50 · ImageNet-1k val (default training, CoTNet-50) | top-1 accuracy: 79.0 -> 79.2 | Improved |
| replace 3x3 convs with CoT blocks across all stages | ResNet-50 · ImageNet-1k val (default training) | top-1 accuracy: 77.3 -> 79.2 | Improved |
| use CoT in res4 and res5 on SE-ResNetD-50 (ablation) | SE-ResNetD-50 · ImageNet-1k val (default training) | top-1 accuracy: 79.1 -> 79.8 | Improved |
| add exponential moving average | CoTNet-50 · ImageNet-1k val (default training) | top-1 accuracy: 79.2 -> 79.8 | Improved |
| use advanced training (350 epochs, RandAugment, mixup, dropout) | CoTNet-50 · ImageNet-1k val | top-1 accuracy: 79.2 -> 81.3 | Improved |
| use ResNeXt-style CoTNeXt instead of CoTNet | CoTNet-50 · ImageNet-1k val (default training) | top-1 accuracy: 79.2 -> 79.5 | Improved |
| use CoTNet backbone instead of ResNet backbone | Faster-RCNN · COCO val Faster-RCNN | AP: 39.34 -> 43.50 | Improved |
| use CoTNeXt-101 backbone instead of ResNeSt-101 | Cascade-RCNN · COCO val Cascade-RCNN | AP: 47.51 -> 49.02 | Improved |
Closest papers by meaning
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.90)
- Container: Context Aggregation Networks (NeurIPS 2021 · similarity 0.88)
- Feature Pyramid Transformer (ECCV 2020 · similarity 0.87)
- Dynamic Graph Message Passing Networks for Visual Recognition (TPAMI 2022 · similarity 0.87)
- UniFormer: Unifying Convolution and Self-Attention for Visual Recognition (TPAMI 2023 · similarity 0.87)
- A Survey on Vision Transformer (TPAMI 2022 · 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.
