Research map / Vision-language and multimodal / Human and embodied modeling
Top-Down Visual Attention From Analysis by Synthesis
Baifeng Shi, Trevor Darrell, Xin Wang
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 prior-conditioned top-down feedback to ViT-B | Vision-language model with ViT-B visual backbone · VQAv2 test-std | accuracy: 67.92 -> 68.78 | Improved |
| add prior-conditioned top-down feedback to ViT-B | Vision-language model with ViT-B visual backbone · Flickr30K zero-shot | IR@1: 42.40 -> 45.28 | Improved |
| replace top-down feedback with PerceiverIO token reweighting (ablation) | ViT-B vision-language model · VQAv2 test-dev | accuracy: 67.89 -> 67.87 | No clear effect |
| remove prior-based token reweighting but retain feedback (ablation) | ViT-B vision-language model · VQAv2 test-dev | accuracy: 67.89 -> 67.99 | Improved |
| use masked value-token reweighting instead of additive top-down input (ablation) | ViT-B vision-language model · VQAv2 test-dev | accuracy: 67.89 -> 67.53 | Hurt |
| add AbS feedback to ViT-Ti | ViT-Ti · ImageNet-1K | accuracy: 72.5 -> 74.1 | Improved |
| add AbS feedback to ViT-Ti | ViT-Ti · ImageNet-C | mCE: 71.1 -> 66.7 | Improved |
| add AbS feedback to ViT-S | ViT-S · ImageNet-A | accuracy: 19.2 -> 24.3 | Improved |
| add AbS feedback to ViT-B | ViT-B · ImageNet-SK | accuracy: 43.3 -> 42.9 | Hurt |
| add AbSViT-B as the segmentation backbone | ViT-B segmentation backbone · PASCAL VOC | segmentation score: 80.1 -> 81.3 | Improved |
| add top-down input to query, key, and value rather than value only (ablation) | AbSViT · ImageNet-1K | accuracy: 74.1 -> 73.3 | Hurt |
| add top-down input to query, key, and value rather than value only (ablation) | AbSViT · ImageNet-C | mCE: 66.7 -> 68.0 | Hurt |
Closest papers by meaning
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 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)
- Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention (AAAI 2024 · similarity 0.86)
- See What You Are Told: Visual Attention Sink in Large Multimodal Models (ICLR 2025 · 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.
