Research map / Architectures and efficiency / CNN design
FOVI: A biologically-inspired foveated interface for deep vision models
Nicholas Blauch, George A. Alvarez, Talia Konkle
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 |
|---|---|---|---|
| use foveated kNN-convolution patch embedding (a=2.79) instead of uniform downsampling to 64x64 | DINOv3 ViT-S+ @ 64, 1 fixation · ImageNet-1K val | top-1 accuracy: 0.693 -> 0.700 | Improved |
| use foveated sampling instead of log-polar sampling | DINOv3 ViT-S+ @ 64, 1 fixation · ImageNet-1K val | top-1 accuracy: 0.643 -> 0.700 | Improved |
| increase foveation from weak (a=60.94) to a=2.79 (ablation) | DINOv3 ViT-S+ @ 64, 1 fixation · ImageNet-1K val | top-1 accuracy: 0.693 -> 0.700 | Improved |
| increase foveation from weak (a=60.94) to a=2.79 (ablation) | DINOv3 ViT-S+ @ 64, 3 fixations · ImageNet-1K val | top-1 accuracy: 0.725 -> 0.735 | Improved |
| add two more fixations (1 -> 3) | FOVI-ViT-H+ @ 64 · ImageNet-1K val | top-1 accuracy: 0.844 -> 0.853 | Improved |
| add two more fixations (1 -> 3) | FOVI-ViT-S+ @ 64 · ImageNet-1K val | top-1 accuracy: 0.700 -> 0.735 | Improved |
| replace uniform @ 224 with FOVI @ 64 (1/16 pixels) | DINOv3 ViT-H+ · ImageNet-1K val | GFLOPS per image: 172.39 -> 58.43 | Improved |
| replace uniform @ 224 with FOVI @ 64 (1/16 pixels) | DINOv3 ViT-H+ · ImageNet-1K val | top-1 accuracy: 0.871 -> 0.844 | Hurt |
| raise sensor resolution from 64 to 224 pixels per side with the same 64 patches (ablation) | FOVI-ViT-S+ (1 fixation) · ImageNet-1K val | top-1 accuracy: 0.700 -> 0.737 | Improved |
| LoRA on patch embedding and first half of network instead of frozen adaptation (ablation) | DINOv3 on foveated input · ImageNet-100 | accuracy: outperforms frozen adaptation by ~30% | Improved |
| LoRA instead of full fine-tuning (ablation) | DINOv3 on foveated input · ImageNet-100 | accuracy: outperforms full fine-tuning by ~10% | Improved |
| adapt first half of the network instead of only late layers (ablation) | DINOv3 on foveated input (LoRA) · ImageNet-100 | accuracy: outperforms late-only adaptation by ~15% | Improved |
Closest papers by meaning
- Seeing More with Less: Human-like Representations in Vision Models (CVPR 2025 · similarity 0.89)
- CanViT: Toward Active-Vision Foundation Models (NeurIPS 2026 · similarity 0.87)
- TransNeXt: Robust Foveal Visual Perception for Vision Transformers (CVPR 2024 · similarity 0.86)
- The All-Seeing Project: Towards Panoptic Visual Recognition and Understanding of the Open World (ICLR 2024 · similarity 0.85)
- LLMind: Bio-inspired Training-free Adaptive Visual Representations for Vision-Language Models (CVPR 2026 · similarity 0.85)
- Foveated Instance Segmentation (CVPR 2025 · similarity 0.84)
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.
