Research map / Architectures and efficiency / Attention and transformers
Vision Transformer With Progressive Sampling
Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip H.S. Torr, Wayne Zhang, Dahua Lin
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 |
|---|---|---|---|
| increase sampling points per axis from n=10 to n=18 (ablation) | PS-ViT-B · ImageNet-1k | top-1 accuracy (%): 80.6 -> 82.3 | Improved |
| increase iterations of progressive sampling from N=1 (fixed locations) to N=4 (ablation) | PS-ViT-B/14 · ImageNet-1k | top-1 accuracy (%): 80.6 -> 81.7 | Improved |
| increase iterations from N=4 to N=8 (ablation) | PS-ViT-B/14 · ImageNet-1k | top-1 accuracy (%): 81.7 -> 81.9 | No clear effect |
| increase iterations beyond N=8 (fewer transformer layers remain) (ablation) | PS-ViT-B/14 · ImageNet-1k | top-1 accuracy (%): 81.9 -> 81.6 | Hurt |
| add progressive sampling module in place of ViT hard patch splitting (ablation) | ViT with same configuration · ImageNet-1k | top-1 accuracy (%): 78.4 -> 81.7 | Improved |
| share weights across progressive sampling iterations (ablation) | PS-ViT-B/14 · ImageNet-1k | top-1 accuracy (%): 81.7 -> 81.5 | Hurt |
| share weights across progressive sampling iterations (ablation) | PS-ViT-Ti/14 · ImageNet-1k | top-1 accuracy (%): 75.6 -> 74.1 | Hurt |
| share weights across progressive sampling iterations (ablation) | PS-ViT-B/10 · ImageNet-1k | top-1 accuracy (%): 80.6 -> 80.0 | Hurt |
| use PS-ViT-B/14 instead of DeiT-B for transfer | PS-ViT-B/14 · Flowers-102 | top-1 accuracy (%): 98.4 -> 98.8 | Improved |
| use PS-ViT-B/14 instead of DeiT-B for transfer | PS-ViT-B/14 · Stanford Cars | top-1 accuracy (%): 92.1 -> 92.9 | Improved |
| use PS-ViT-B/14 instead of DeiT-B for transfer | PS-ViT-B/14 · CIFAR-100 | top-1 accuracy (%): 90.8 -> 90.8 | No clear effect |
| use PS-ViT-B/14 instead of DeiT-B at similar accuracy | PS-ViT-B/14 · Speed (V100) | images per second: 316.1 -> 765.6 | Improved |
Closest papers by meaning
- A Survey on Efficient Vision Transformers: Algorithms, Techniques, and Performance Benchmarking (TPAMI 2024 · similarity 0.89)
- Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · similarity 0.89)
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021 · similarity 0.89)
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.87)
- Sub-token ViT Embedding via Stochastic Resonance Transformers (ICML 2024 · similarity 0.87)
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.
