Research map / Detection and segmentation / CNN features
Bottleneck Transformers for Visual Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the final three spatial convolutions with global self-attention | R50 backbone · COCO val, 36 epochs, jitter [0.8, 1.25] | mask AP: 37.7 -> 38.9 | Helped |
| Replace the final three spatial convolutions with global self-attention | R50 backbone · COCO val, 12 epochs | mask AP: 35.0 -> 35.3 | No clear effect |
| Use longer training with the default scale jitter | BoT50 · COCO val, 72 vs 36 epochs, jitter [0.8, 1.25] | box AP gain over R50: +1.5 -> +0.9 | Hurt |
| Increase multi-scale jitter to [0.1, 2.0] | BoT50 relative to R50 · COCO val, 72 epochs | mask AP gain: +0.8 -> +1.5 | Helped |
| Use content-content attention interaction qkᵀ (ablation) | BoT50 · COCO val, 36 epochs | box AP: +0.6 | Helped |
| Add relative content-position interaction qrᵀ (ablation) | BoT50 · COCO val, 36 epochs | box AP: +1.0 | Helped |
| Combine content-content and relative content-position interactions (ablation) | BoT50 · COCO val, 36 epochs | box AP: +1.5 | Helped |
| Use absolute rather than relative position encodings (ablation) | BoT50 · COCO val, 36 epochs | box AP: 43.6 -> 42.5 | Hurt |
| Insert one BoT block in c4 instead of a Non-Local block (ablation) | R50 · COCO val, 36 epochs | box AP: 43.1 -> 43.7 | Helped |
| Use stride 1 in all final-group attention layers (ablation) | BoTNet-S1-50 · ImageNet-1K val, 100 epochs | top-1 accuracy: 76.8 -> 77.7 | Helped |
| Use the improved training setting with longer training and stronger augmentation | BoT50 · ImageNet-1K val, 200 epochs | top-1 accuracy: +0.6% | Helped |
| Use the improved training setting with longer training and stronger augmentation | BoT-S1-50 · ImageNet-1K val, 200 epochs | top-1 accuracy: +1.4% | Helped |
Closest papers by meaning
- Scaling Local Self-Attention for Parameter Efficient Visual Backbones (CVPR 2021 · similarity 0.88)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.88)
- Exploring Self-Attention for Image Recognition (CVPR 2020 · similarity 0.88)
- Stand-Alone Self-Attention in Vision Models (NeurIPS 2019 · similarity 0.88)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.87)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · 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.