Research map / Detection and segmentation / CNN features
DETR Does Not Need Multi-Scale or Locality Design
Yutong Lin, Yuhui Yuan, Zheng Zhang, Chen Li, Nanning Zheng, Han Hu
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 BoxRPB to cross-attention (ablation) | plain DETR (Swin-S) · COCO val | AP: 37.2 -> 46.1 | Improved |
| add MIM (SimMIM) backbone pre-training (ablation) | plain DETR (Swin-S) · COCO val | AP: 37.2 -> 44.6 | Improved |
| add MIM pre-training on top of BoxRPB (ablation) | plain DETR (Swin-S) · COCO val | AP: 46.1 -> 48.7 | Improved |
| add box re-parameterization for regression (ablation) | plain DETR with BoxRPB and MIM · COCO val | AP: 48.7 -> 50.9 | Improved |
| replace cross-entropy with focal loss for classification (ablation) | plain DETR baseline · COCO val | AP: 23.1 -> 31.6 | Improved |
| add hybrid matching (ablation) | plain DETR baseline · COCO val | AP: 35.2 -> 37.2 | Improved |
| use axial decomposition of BoxRPB (ablation) | BoxRPB · COCO val | AP and cost: AP 50.8 -> 50.9, GFLOPs 265.4 -> 5.8 | Improved |
| use two box corners instead of box center only for BoxRPB (ablation) | BoxRPB · COCO val | AP: 48.0 -> 50.9 | Improved |
| use MIM pre-training to drop multi-scale backbone features (C5 only) (ablation) | plain DETR · COCO val | AP: supervised: 47.0 -> 46.4; MIM: 50.3 -> 50.2 | Improved |
| use multi-scale decoder feature maps instead of single-scale P4 (ablation) | plain DETR with MIM · COCO val | AP: 50.9 -> 51.1 | No clear effect |
| use global BoxRPB cross-attention instead of local cross-attention (deformable, RoIAlign, RoI sampling, box mask) (ablation) | plain DETR · COCO val | AP: 50.2 (deformable) -> 50.9 | Improved |
| replace DAB cross-attention with BoxRPB in DAB official code (ablation) | DAB-DETR · COCO val | mAP: +1.8 mAP | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Deformable DETR — Deformable DETR: Deformable Transformers for End-to-End Object Detection (ICLR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Decoupled DETR: Spatially Disentangling Localization and Classification for Improved End-to-End Object Detection (ICCV 2023 · similarity 0.91)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.89)
- Less is More: Focus Attention for Efficient DETR (ICCV 2023 · similarity 0.89)
- CF-DETR: Coarse-to-Fine Transformers for End-to-End Object Detection (AAAI 2022 · similarity 0.89)
- V-DETR: DETR with Vertex Relative Position Encoding for 3D Object Detection (ICLR 2024 · similarity 0.87)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 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.
