Research map / Detection and segmentation / CNN features
Detection Transformer with Stable Matching
Shilong Liu, Tianhe Ren, Jiayu Chen, Zhaoyang Zeng, Hao Zhang, Feng Li, Hongyang Li, Jun Huang, Hang Su, Jun Zhu, Lei Zhang
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 position-supervised loss (IoU-based target for positive classification scores) (ablation) | DINO · COCO val2017 (R50, 12 epochs) | AP: 49.2 -> 49.8 | Improved |
| add position-modulated matching cost (ablation) | DINO + position-supervised loss · COCO val2017 (R50, 12 epochs) | AP: 49.8 -> 50.2 | Improved |
| add dense memory fusion (ablation) | DINO + PSL + PMC · COCO val2017 (R50, 12 epochs) | AP: 50.2 -> 50.4 | Improved |
| use simple memory fusion instead of dense fusion (ablation) | DINO + PSL + PMC · COCO val2017 (R50, 12 epochs) | AP: 50.2 -> 50.2 | No clear effect |
| add classification probability to the positive-example target (s*p) (ablation) | position-supervised loss · COCO val2017 (R50, 12 epochs) | AP: 49.2 -> 26.4 | Hurt |
| use convex target s^2 instead of concave s^0.5 (ablation) | position-supervised loss · COCO val2017 (R50, 12 epochs) | AP: 49.3 -> 49.6 | Improved |
| increase classification loss weight to 20.0 (ablation) | position-supervised loss · COCO val2017 (R50, 12 epochs) | AP: 50.2 -> 48.3 | Hurt |
| use position-modulated cost with f2=s^2 instead of s^0.5 (ablation) | position-modulated cost · COCO val2017 (R50, 12 epochs) | AP: 50.2 -> 48.8 | Hurt |
| apply stable matching to other DETR variants | Deformable-DETR, DAB-Deformable-DETR, H-DETR · COCO val2017 (R50, 12 epochs) | AP: +1.3, +1.0, +0.6 | Improved |
| apply stable matching to MaskDINO | MaskDINO · COCO val2017 instance segmentation | mask AP / box AP: 41.4 -> 42.1 mask AP; 45.7 -> 47.0 box AP | Improved |
| apply stable matching to DINO with Swin-L at 12 epochs | DINO · COCO val2017 (Swin-L) | AP: 56.8 -> 57.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MaskDINO — Mask DINO: Towards a Unified Transformer-Based Framework for Object Detection and Segmentation (CVPR 2023) · better on 2 of 2 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
- Accelerating DETR Convergence via Semantic-Aligned Matching (CVPR 2022 · similarity 0.88)
- CF-DETR: Coarse-to-Fine Transformers for End-to-End Object Detection (AAAI 2022 · similarity 0.87)
- DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR (ICLR 2022 · similarity 0.87)
- KD-DETR: Knowledge Distillation for Detection Transformer with Consistent Distillation Points Sampling (CVPR 2024 · similarity 0.87)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.86)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.85)
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.
