Research map / Detection and segmentation / CNN features
A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation
Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang, Denny Zhou
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 |
|---|---|---|---|
| remove spatial downsampling, multi-scale features and doubled channels (plain ViT) (ablation) | ViT, 72M params · COCO val | box AP: vanilla ViT performs best across FLOPs range | Improved |
| add multi-scale features to spatial downsampling (ablation) | ViT · COCO val | box AP | Improved |
| use 18 blocks instead of 12 or 40 (ablation) | UViT · COCO val | box AP | Improved |
| use 896x896 input instead of 1024x1024 or smaller (ablation) | UViT · COCO val | box AP | Improved |
| use window scale 1/2 instead of global attention in all blocks (ablation) | UViT-B · COCO val | box AP: 52.4 -> 52.3 (GFLOPs 2961.9 -> 1298.7) | No clear effect |
| over-shrink early windows to 1/16 (ablation) | UViT-B · COCO val | box AP: 52.5 -> 52.0 | Hurt |
| use local windows (1/2) in all deep layers instead of global windows at end (ablation) | UViT-B · COCO val | box AP: 52.5 -> 52.2 | Hurt |
| progressive windows 4^-1 x14, 2^-1 x2, 1 x2 (UViT+) instead of constant 1/2 (ablation) | UViT-B · COCO val | box AP: 52.3 -> 52.5 (GFLOPs 1298.7 -> 1160.1) | Improved |
| add self-training with unlabeled ImageNet-1K | UViT-B+ · COCO val | box AP: 52.5 -> 53.9 | Improved |
| add self-training with unlabeled ImageNet-1K | UViT-B+ · COCO val | mask AP: 44.8 -> 46.1 | Improved |
| use 32 blocks instead of shallower (ablation) | UViT · Pascal VOC | mIoU | Improved |
Closest papers by meaning
- Vision Transformers: From Semantic Segmentation to Dense Prediction (IJCV 2024 · similarity 0.91)
- A Survey on Efficient Vision Transformers: Algorithms, Techniques, and Performance Benchmarking (TPAMI 2024 · similarity 0.91)
- Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection (ICCV 2023 · similarity 0.90)
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.90)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.90)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · similarity 0.89)
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.
