Research map / Architectures and efficiency / Attention and transformers
Iterative Patch Selection for High-Resolution Image Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Use IPS patch selection instead of random patch selection | Transformer on Swedish traffic signs · Swedish traffic signs test | accuracy: 50.7 ± 2.6 -> 98.6 ± 0.5 | Helped |
| Use IPS patch selection instead of differential patch selection | Transformer on Swedish traffic signs · Swedish traffic signs test | accuracy: 94.0 ± 2.2 -> 98.6 ± 0.5 | Helped |
| Use IPS cross-attention instead of DeepMIL aggregation | Traffic sign recognition with all 192 patches · Swedish traffic signs test | accuracy: 96.2 ± 1.6 -> 98.4 ± 0.5 | Helped |
| Downscale the input image by a factor of 3 | CNN on Swedish traffic signs · Swedish traffic signs test | accuracy: 84.1 ± 4.7 -> 76.3 ± 0.9 | Hurt |
| Use IPS to process all WSI patches instead of a 70k-patch DeepMIL sample | CAMELYON16 classification · CAMELYON16 test | AUC: 94.5 ± 1.1 -> 98.1 ± 0.3 | Helped |
| Reduce selected patches from 5k to 100 (ablation) | IPS Transformer on CAMELYON16 · CAMELYON16 test | AUC: 98.1 ± 0.3 -> 97.3 ± 0.6 | Hurt |
| Use IPS instead of random patch selection | Transformer on megapixel MNIST · Megapixel MNIST (N=3,481), multi-label | accuracy: 0.9 ± 0.3 -> 96.1 ± 0.7 | Helped |
| Use overlapping patches (N=3,481) instead of non-overlapping patches (N=900) (ablation) | IPS Transformer (M=100) · Megapixel MNIST, multi-label | accuracy: 84.2 ± 2.0 -> 96.1 ± 0.7 | Helped |
| Increase selected patches from 5 to 100 (ablation) | IPS Transformer with overlapping patches · Megapixel MNIST, multi-label | accuracy: 90.2 ± 1.0 -> 96.1 ± 0.7 | Helped |
| Increase image side length using lazy loading (ablation) | IPS Transformer on megapixel MNIST · Megapixel MNIST, image sizes from 1k to 10k pixels per side | peak GPU memory: 1.7 GB | No clear effect |
| Increase image side length beyond 8k pixels (ablation) | IPS Transformer on megapixel MNIST · Megapixel MNIST, majority task | accuracy | Hurt |
| Increase I, the patches processed per selection iteration (ablation) | IPS on megapixel MNIST · Megapixel MNIST | runtime | Mixed |
Closest papers by meaning
- No Pains, More Gains: Recycling Sub-Salient Patches for Efficient High-Resolution Image Recognition (CVPR 2025 · similarity 0.86)
- Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More (ICML 2025 · similarity 0.84)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.83)
- Patch-Aware Sample Selection for Efficient Masked Image Modeling (AAAI 2024 · similarity 0.83)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.83)
- Ultra-High-Definition Dynamic Multi-Exposure Image Fusion via Infinite Pixel Learning (AAAI 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.