Research map / Architectures and efficiency / Attention and transformers
Glance-and-Gaze Vision Transformer
Qihang Yu, Yingda Xia, Yutong Bai, Yongyi Lu, Alan Yuille, Wei Shen
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 Gaze depthwise-conv branch to Glance attention (ablation) | GG-T · ImageNet-1K 100-epoch | top-1 accuracy: 77.21% -> 80.28% | Improved |
| add Glance attention branch to Gaze conv (ablation) | GG-T · ImageNet-1K 100-epoch | top-1 accuracy: 76.76% -> 80.28% | Improved |
| implement Gaze as depthwise conv instead of window self-attention (ablation) | GG-T · ImageNet-1K 100-epoch | top-1 accuracy: 79.07% -> 80.28% | Improved |
| replace W&SW-MSA with Glance+Gaze (Conv) (ablation) | Swin-T · ImageNet-1K 100-epoch | top-1 accuracy: 78.50% -> 80.28% | Improved |
| replace W&SW-MSA with full MSA in stages 3 and 4 (ablation) | Swin-T · ImageNet-1K 100-epoch | top-1 accuracy: 78.50% -> 79.79% | Improved |
| use adaptive gaze kernels (9,5,3,3) instead of fixed 3x3 (ablation) | GG-T · ImageNet-1K 100-epoch | top-1 accuracy: 80.28% -> 80.38% | No clear effect |
| use fixed 5x5 gaze kernels instead of 3x3 (ablation) | GG-T · ImageNet-1K 100-epoch | top-1 accuracy: 80.28% -> 80.31% | No clear effect |
| replace MSA with GG-MSA | DeiT-T · ImageNet-1K | top-1 accuracy: 72.2% -> 73.8% | Improved |
| replace MSA with GG-MSA | DeiT-S · ImageNet-1K | top-1 accuracy: 79.9% -> 80.5% | Improved |
| replace Swin window attention with Glance-and-Gaze | Swin-T/S · ImageNet-1K | top-1 accuracy: 81.2 -> 82.0 (T), 83.2 -> 83.4 (S) | Improved |
| replace Swin window attention with Glance-and-Gaze | Swin-S on COCO Mask R-CNN · COCO val2017 | mask AP: 41.4 -> 41.3 (mask AP), 45.4 -> 45.7 (box AP) | Mixed |
Closest papers by meaning
- MaxViT: Multi-axis Vision Transformer (ECCV 2022 · similarity 0.88)
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.88)
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 2024 · similarity 0.87)
- TransGOP: Transformer-Based Gaze Object Prediction (AAAI 2024 · similarity 0.87)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.86)
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · 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.
