Research map / Vision-language and multimodal / Human and embodied modeling
Peripheral Vision Transformer
Juhong Min, Yucheng Zhao, Chong Luo, Minsu Cho
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 position-based peripheral attention Phi_p (ablation) | PerViT-T · ImageNet-1K val | Top-1 accuracy: 77.3 -> 78.8 | Improved |
| add position-based peripheral attention Phi_p (ablation) | PerViT-S · ImageNet-1K val | Top-1 accuracy: 81.1 -> 82.1 | Improved |
| add position-based peripheral attention Phi_p (ablation) | PerViT-M · ImageNet-1K val | Top-1 accuracy: 81.9 -> 82.9 | Improved |
| add content-based attention Phi_c (ablation) | PerViT-T · ImageNet-1K val | Top-1 accuracy: 76.8 -> 78.8 | Improved |
| add convolutional stem (ablation) | PerViT-T · ImageNet-1K val | Top-1 accuracy: 77.8 -> 78.8 | Improved |
| add convolutional position encoding (CPE) (ablation) | PerViT-T · ImageNet-1K val | Top-1 accuracy: 78.1 -> 78.8 | Improved |
| add C-Stem and CPE (ablation) | PerViT without Phi_p · ImageNet-1K val | Top-1 accuracy: T: 72.2 -> 77.3; S: 79.8 -> 81.1; M: 81.8 -> 81.9 | Improved |
| peripheral initialization instead of convolutional initialization (ablation) | PerViT-T Phi_p · ImageNet-1K val | Top-1 accuracy: 78.6 -> 78.8 | Improved |
| peripheral initialization instead of random initialization (ablation) | PerViT-T Phi_p · ImageNet-1K val | Top-1 accuracy: 78.5 -> 78.8 | Improved |
| add peripheral projection (neighbors) and multi-layer jointly (ablation) | Phi_p with Euclidean distance · ImageNet-1K val | Top-1 accuracy: 78.0 -> 78.8 | Improved |
| add multi-layer projection alone (no peripheral projection) (ablation) | Phi_p with Euclidean distance · ImageNet-1K val | Top-1 accuracy: 78.0 -> 77.9 | No clear effect |
| Euclidean distance embedding instead of learnable vectors (ablation) | Phi_p · ImageNet-1K val | Top-1 accuracy: 77.8 -> 78.8 | Improved |
Closest papers by meaning
- COCO-Periph: Bridging the Gap Between Human and Machine Perception in the Periphery (ICLR 2024 · similarity 0.89)
- RegionViT: Regional-to-Local Attention for Vision Transformers (ICLR 2022 · similarity 0.85)
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.85)
- TransNeXt: Robust Foveal Visual Perception for Vision Transformers (CVPR 2024 · similarity 0.84)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.83)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.83)
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.
