Research map / Detection and segmentation / CNN features
Exploring Plain Vision Transformer Backbones for Object Detection
Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He
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 simple feature pyramid (ablation) | ViTDet ViT-L · COCO val | AP box: 51.2 -> 54.6 | Improved |
| replace simple pyramid with FPN 4-stage (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 54.4 | No clear effect |
| replace simple pyramid with average-pooling-only pyramid (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 54.5 | No clear effect |
| add 4 global propagation blocks to window attention (ablation) | ViTDet ViT-L · COCO val | AP box: 52.9 -> 54.6 | Improved |
| add 4 convolutional propagation blocks (ablation) | ViTDet ViT-L · COCO val | AP box: 52.9 -> 54.8 | Improved |
| use shifted windows instead of propagation blocks (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 54.0 | Hurt |
| place propagation in first 4 blocks only (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 52.9 | Hurt |
| place propagation in last 4 blocks instead of evenly (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 54.3 | No clear effect |
| use 24 global blocks instead of 4 (ablation) | ViTDet ViT-L · COCO val | AP box: 54.6 -> 55.1 | No clear effect |
| use MAE IN-1K pre-training instead of supervised IN-1K (ablation) | ViTDet ViT-L · COCO val | AP box: 49.6 -> 54.6 | Improved |
| use IN-21K supervised instead of IN-1K supervised pre-training (ablation) | ViTDet ViT-L · COCO val | AP box: 49.6 -> 50.6 | No clear effect |
| use IN-1K supervised pre-training instead of random init (ablation) | ViTDet ViT-L · COCO val | AP box: 50.0 -> 49.6 | Hurt |
Closest papers by meaning
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.90)
- Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection (ICCV 2023 · similarity 0.90)
- Efficient Decoder-Free Object Detection with Transformers (ECCV 2022 · similarity 0.90)
- CBNet: A Novel Composite Backbone Network Architecture for Object Detection (AAAI 2020 · similarity 0.89)
- ViDT: An Efficient and Effective Fully Transformer-based Object Detector (ICLR 2022 · similarity 0.89)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 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.
