Research map / Detection and segmentation / CNN features
Should All Proposals Be Treated Equally in Object Detection?
Yunsheng Li, Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Pei Yu, Ying Jin, Lu Yuan, Zicheng Liu, Nuno Vasconcelos
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 complexity loss L_c to the IoU loss (ablation) | DPP · COCO val2017 (ResNet-50) | AP: 41.1 -> 42.2 | Improved |
| add IoU loss L_iou to the complexity loss (ablation) | DPP · COCO val2017 (ResNet-50) | AP: 41.8 -> 42.2 | Improved |
| increase lambda from 10 to 100 (ablation) | DPP · COCO val2017 (ResNet-50) | AP: 42.2 -> 41.7 | Hurt |
| increase lambda from 10 to 300 (ablation) | DPP · COCO val2017 (ResNet-50) | AP: 42.2 -> 41.0 | Hurt |
| use lambda=1 instead of 10 (ablation) | DPP · COCO val2017 (ResNet-50) | AP: 42.1 vs 42.2 | No clear effect |
| add medium operator g1 to g0 (ablation) | DPP last stage · COCO val2017 (ResNet-50) | AP: 41.0 -> 41.7 | Improved |
| add identity operator g2 proposals to g0+g1 (ablation) | DPP last stage · COCO val2017 (ResNet-50) | AP: 41.7 -> 42.2 | Improved |
| replace Sparse R-CNN head by DPP-XL | Sparse R-CNN · COCO val2017 (ResNet-50) | AP / GFLOPS: AP 45.0 -> 45.0, GFLOPS 25 -> 15 | Improved |
| use DPP-S instead of Sparse R-CNN | Sparse R-CNN · COCO val2017 (ResNet-50) | AP: 45.0 -> 40.4 (decrease of 4.6 points, 12.5x less computation) | Hurt |
| replace Sparse R-CNN with DPP-XL on MobileNetV2 | Sparse R-CNN · COCO val2017 (MobileNetV2) | AP: 36.6 -> 36.9 | Improved |
| use DPP-S instead of Sparse R-CNN on MobileNetV2 | Sparse R-CNN · COCO val2017 (MobileNetV2) | AP: 36.6 -> 35.7 | Hurt |
| add more stages of the detection head | DPP · COCO val2017 (ResNet-50) | AP: 15.6, 32.1, 39.3, 41.7, 42.0, 42.2 | Mixed |
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.
- Sparse R-CNN — Sparse R-CNN: End-to-End Object Detection With Learnable Proposals (CVPR 2021) · better on 3 of 4 shared results
- Deformable DETR — Deformable DETR: Deformable Transformers for End-to-End Object Detection (ICLR 2021) · better on 2 of 2 shared results
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
- What Makes for Effective Detection Proposals? (TPAMI 2016 · similarity 0.90)
- Object-Proposal Evaluation Protocol is 'Gameable' (CVPR 2016 · similarity 0.90)
- Dynamic Sparse R-CNN (CVPR 2022 · similarity 0.87)
- Learning Detection With Diverse Proposals (CVPR 2017 · similarity 0.87)
- Proposal-Contrastive Pretraining for Object Detection from Fewer Data (ICLR 2023 · similarity 0.87)
- Toward Scale-Invariance and Position-Sensitive Region Proposal Networks (ECCV 2018 · 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.
