Research map / Detection and segmentation / CNN features
Co-Scale Conv-Attentional Image Transformers
Weijian Xu, Yifan Xu, Tyler Chang, Zhuowen Tu
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 convolutional position encoding (CPE) and relative encoding (CRPE) together to CoaT-Lite Tiny with no position encoding (ablation) | CoaT · ImageNet-1K val (CoaT-Lite Tiny) | Top-1 accuracy (%): 68.8 -> 77.5 | Improved |
| add CRPE to CoaT-Lite Tiny without position encoding (ablation) | CoaT · ImageNet-1K val (CoaT-Lite Tiny) | Top-1 accuracy (%): 68.8 -> 75.0 | Improved |
| add CPE to CoaT-Lite Tiny without position encoding (ablation) | CoaT · ImageNet-1K val (CoaT-Lite Tiny) | Top-1 accuracy (%): 68.8 -> 75.9 | Improved |
| add CRPE on top of CPE (ablation) | CoaT · ImageNet-1K val (CoaT-Lite Tiny) | Top-1 accuracy (%): 75.9 -> 77.5 | Improved |
| add CPE on top of CRPE (ablation) | CoaT · ImageNet-1K val (CoaT-Lite Tiny) | Top-1 accuracy (%): 75.0 -> 77.5 | Improved |
| add co-scale with attention via feature interpolation to CoaT Tiny (ablation) | CoaT · ImageNet-1K val | Top-1 accuracy (%): 77.8 -> 78.3 | Improved |
| add co-scale with attention via feature interpolation to CoaT Tiny (detection) (ablation) | CoaT · COCO Mask R-CNN FPN 1x box AP | AP: 41.6 -> 42.5 | Improved |
| add co-scale with direct cross-layer attention to CoaT Tiny (ablation) | CoaT · ImageNet-1K val | Top-1 accuracy (%): 77.8 -> 77.0 | Hurt |
| add co-scale with direct cross-layer attention to CoaT Tiny (detection) (ablation) | CoaT · COCO Mask R-CNN FPN 1x box AP | AP: 41.6 -> 42.1 | Improved |
| use feature interpolation instead of direct cross-layer attention for co-scale (ablation) | CoaT · ImageNet-1K val | Top-1 accuracy (%): 77.0 -> 78.3 | Improved |
| use CoaT Small instead of ResNet-50 backbone in Deformable DETR | CoaT · COCO Deformable DETR AP | AP: 44.5 -> 48.4 | Improved |
| add parallel group with co-scale (CoaT Small vs CoaT-Lite Small) on classification (ablation) | CoaT · ImageNet-1K val | Top-1 accuracy (%): 81.9 -> 82.1 | Improved |
Closest papers by meaning
- CoAtNet: Marrying Convolution and Attention for All Data Sizes (NeurIPS 2021 · similarity 0.89)
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021 · similarity 0.87)
- Contextual Transformer Networks for Visual Recognition (TPAMI 2022 · similarity 0.87)
- Less Is More: Pay Less Attention in Vision Transformers (AAAI 2022 · similarity 0.86)
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.86)
- UniFormer: Unifying Convolution and Self-Attention for Visual Recognition (TPAMI 2023 · 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.
