Research map / Detection and segmentation / CNN features
Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction Without Convolutions
Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao
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 progressive shrinking pyramid structure instead of columnar ViT-Small/32 (ablation) | RetinaNet · COCO val2017 | AP: 31.7 -> 40.4 | Improved |
| use fine-grained 4x4 patches in a columnar ViT (ablation) | ViT-Small/4 RetinaNet · COCO val2017 | AP | Hurt |
| go deeper (PVT-Medium) instead of wider (x1.4 hidden dims) (ablation) | PVT-Small · COCO val2017 RetinaNet 1x | AP: 40.8 -> 41.9 | Improved |
| go deeper instead of wider (ablation) | PVT-Small · ImageNet-1k val | top-1 error (%): 19.3 -> 18.8 | Improved |
| add ImageNet pre-trained weights (ablation) | RetinaNet-PVT-Small · COCO val2017 | AP: gap of 13.8 (1x) and 8.4 (3x + MS) | Improved |
| replace CNN with non-local (ResNet50+GC r4) by PVT-Small (ablation) | Mask R-CNN 1x · COCO val2017 | mask AP: 36.2 -> 37.8 | Improved |
| reduce input shorter side from 800 to 640 | RetinaNet PVT-Small · COCO val2017 | AP: 40.4 -> 38.7; 76.9ms -> 51.7ms | Mixed |
| replace ResNet50 with PVT-Small backbone | RetinaNet 1x · COCO val2017 | AP: 36.3 -> 40.4 | Improved |
| replace ResNet50 with PVT-Small backbone | DETR 50 epochs · COCO val2017 | AP: 32.3 -> 34.7 | Improved |
| replace ResNet50-d16 with PVT-Small in Trans2Seg | Trans2Seg · ADE20K val | mIoU (%): 39.7 -> 42.6 | Improved |
| replace ResNet18/50/101 with PVT-Tiny/Small/Medium | Semantic FPN · ADE20K val | mIoU (%): 32.9 -> 35.7; 36.7 -> 39.8 | Improved |
| replace ResNet50 with PVT-Small | classification · ImageNet-1k val | top-1 error (%): 21.5 -> 20.2 | Improved |
Closest papers by meaning
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.91)
- Vision Transformers for Dense Prediction (ICCV 2021 · similarity 0.90)
- Vision Transformers: From Semantic Segmentation to Dense Prediction (IJCV 2024 · similarity 0.90)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.90)
- CMT: Convolutional Neural Networks Meet Vision Transformers (CVPR 2022 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.88)
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.
