Research map / Detection and segmentation / CNN features
GPViT: A High Resolution Non-Hierarchical Vision Transformer with Group Propagation
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 |
|---|---|---|---|
| add GP Block in place of window attention (ablation) | window-attention GPViT (L1) · ImageNet-1K | Top-1 Acc: 78.1 -> 79.8 | Helped |
| add GP Block on top of LePE attention (ablation) | LePE-attention GPViT (L1) · ImageNet-1K | Top-1 Acc: 79.5 -> 80.5 | Helped |
| add GP Block on top of window attention (ablation) | window-attention GPViT (L1) · ADE20K SegFormer | mIoU: 41.7 -> 45.5 | Helped |
| add GP Block on top of LePE attention (ablation) | LePE-attention GPViT (L1) · ADE20K SegFormer | mIoU: 46.2 -> 46.9 | Helped |
| replace local attention with convolution propagation blocks (ablation) | GPViT (window and LePE attention) · ImageNet-1K | Top-1 Acc: 78.2 -> 75.8 (window); 79.5 -> 77.7 (LePE) | Hurt |
| replace GP Block with global attention (ablation) | GPViT (LePE attention) · ImageNet-1K | Top-1 Acc: 80.4 vs 80.5 for GP Block at ~10x FLOPs (62G vs 34G ADE20K) | Mixed |
| use shifted window attention instead of GP Block (ablation) | window-attention GPViT · ADE20K SegFormer | mIoU: 41.7 -> 40.7 | Hurt |
| remove group propagation (grouping and ungrouping only) (ablation) | GP Block · ImageNet-1K | Top-1 Acc: 80.5 -> 79.8 | Hurt |
| use self-attention instead of MLPMixer for group propagation (ablation) | GP Block · ImageNet-1K | Top-1 Acc: 80.5 -> 80.7 at higher FLOPs (6.2G vs 5.8G) | Helped |
| increase patch resolution (patch 16 to 8) with global self-attention (ablation) | ViT-D216 · ImageNet-1K | Top-1 Acc: 77.4 -> 79.2 | Helped |
| GPViT-L1 vs Swin-B backbone in Mask R-CNN 1x | GPViT-L1 · COCO Mask R-CNN 1x | AP^bb: 45.5 -> 48.1 | Helped |
| GPViT-L3 vs Swin-B on ADE20K UperNet | GPViT-L3 · ADE20K UperNet | mIoU: 49.7 -> 51.7 | Helped |
Closest papers by meaning
- GroupViT: Semantic Segmentation Emerges From Text Supervision (CVPR 2022 · similarity 0.91)
- Green Hierarchical Vision Transformer for Masked Image Modeling (NeurIPS 2022 · similarity 0.88)
- AdaViT: Adaptive Vision Transformers for Efficient Image Recognition (CVPR 2022 · similarity 0.87)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · similarity 0.87)
- Perceptual Group Tokenizer: Building Perception with Iterative Grouping (ICLR 2024 · similarity 0.86)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 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.