Research map / Detection and segmentation / CNN features
Improving Object Localization With Fitness NMS and Bounded IoU Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train with matching IoU 0.7 instead of 0.5 (ablation) | DeNet-34 · MSCOCO test-dev | MAP@0.5:0.95: 29.5 -> 30.7; +1.2% | Helped |
| Add independent fitness prediction to NMS scoring | DeNet-34 (skip) · VOC 2007 test, Ω_test=0.9 | MAP: 18.4 -> 25.3 | Helped |
| Add joint class-and-fitness prediction to NMS scoring | DeNet-34 (skip) · VOC 2007 test, Ω_test=0.9 | MAP: 18.4 -> 27.9 | Helped |
| Use independent rather than joint fitness prediction (ablation) | DeNet models · MSCOCO test-dev, Ω_test=0.5 | MAP: DeNet-34 S: 47.9 -> 47.7; DeNet-34 W: 48.9 -> 48.5; DeNet-101 W: 53.4 -> 52.9 | Hurt |
| Use joint rather than independent fitness prediction (ablation) | DeNet-101 wide · MSCOCO test-dev, Ω_test=0.5:0.95 | MAP: 35.2 -> 36.5 | Helped |
| Add joint Fitness NMS | DeNet models · MSCOCO test-dev, Ω_test=0.75 | MAP: DeNet-34 S: 31.1 -> 33.5; DeNet-34 W: 31.8 -> 34.9; DeNet-101 W: 36.1 -> 39.1 | Helped |
| Combine Gaussian Soft-NMS with joint Fitness NMS | DeNet models · MSCOCO test-dev, MAP@0.5:0.95 | MAP: 31.4 -> 32.2; 32.6 -> 33.4; 36.5 -> 37.2 | Helped |
| Replace R-CNN bounding-box regression loss with Bounded IoU Loss (ablation) | DeNet models with joint Fitness NMS · MSCOCO test-dev, MAP@0.5:0.95 | MAP: 31.4 -> 32.1; 32.6 -> 33.6; 36.5 -> 37.0 | Helped |
| Add corner clustering (m=1) (ablation) | DeNet wide models · MSCOCO test-dev | evaluation rate: 71 Hz -> 79 Hz; 22 Hz -> 24 Hz | Helped |
| Use standard NMS clustering at threshold 0.7 (ablation) | DeNet wide models · MSCOCO test-dev | MAP@0.5:0.95: 33.1 -> 33.7; 36.3 -> 36.9 | Helped |
| Increase input dimension from ×512 to ×768 (ablation) | DeNet-101 wide · MSCOCO test-dev | MAP@0.5:0.95: 36.9 -> 39.5 | Helped |
| Increase input image scale (ablation) | DeNet-101 wide · MSCOCO test-dev, large objects | MAP for large objects: 56.5 -> 54.1 -> 49.4 -> 42.3 -> 34.4 | Mixed |
Closest papers by meaning
- VarifocalNet: An IoU-Aware Dense Object Detector (CVPR 2021 · similarity 0.86)
- Bounding Box Regression With Uncertainty for Accurate Object Detection (CVPR 2019 · similarity 0.86)
- Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection (CVPR 2021 · similarity 0.85)
- D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement (ICLR 2025 · similarity 0.84)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.82)
- Accelerating Non-Maximum Suppression: A Graph Theory Perspective (NeurIPS 2024 · similarity 0.82)
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.