Research map / Detection and segmentation / CNN features
Confluence: A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection
Andrew Shepley, Greg Falzon, Paul Kwan, Ljiljana Branković
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 |
|---|---|---|---|
| replace Greedy NMS and Soft-NMS with Confluence | RetinaNet · MS-COCO mini-val | AP@0.5:0.95: gains in AP of 1.0-1.9% | Improved |
| replace Greedy NMS and Soft-NMS linear with Confluence-NMS (C-NMS) | RetinaNet · MS-COCO mini-val | AP@0.5:0.95: 1.4-2.3% | Improved |
| replace Greedy NMS and Soft-NMS linear with Confluence-NMS | RetinaNet · MS-COCO mini-val | AP at IoU 0.5: 1.6-2.7% | Improved |
| replace Greedy NMS and Soft-NMS with Confluence | RetinaNet · MS-COCO mini-val | AR@100: gains of 1.5-5.3% | Improved |
| use Confluence on classes often occluded or dense (person, truck, zebra, banana) | RetinaNet · MS-COCO mini-val per class | AP: 2.51-9.67% | Improved |
| use Confluence on lone unoccluded classes (laptop, dog, cat, airplane) | RetinaNet · MS-COCO mini-val per class | AP | No clear effect |
| replace Greedy NMS and Soft-NMS with Confluence | Mask-RCNN · CrowdHuman validation | AP@0.5:0.95: gains in AP of 1.2-4% | Improved |
| replace Greedy NMS and Soft-NMS with Confluence | Mask-RCNN · CrowdHuman validation | AR@100: 2.4-7.3% | Improved |
| use Confluence-NMS gaussian instead of Soft-NMS gaussian | Mask-RCNN · CrowdHuman validation | AP@0.5:0.95: 2.9% | Improved |
| use Confluence-NMS gaussian instead of Soft-NMS gaussian | RetinaNet · MS-COCO mini-val | AP@0.5:0.95: 0.4% | Improved |
| use Confluence threshold Ct instead of IoU threshold in the optimal range (ablation) | RetinaNet · MS-COCO mini-val | AP variation over threshold: variation in AP for Ct being 0.3% while variation in AP for IoU is 0.6% | Improved |
| use Confluence even at the optimal IoU threshold of baselines (ablation) | RetinaNet · MS-COCO mini-val | AP@0.5:0.95: approximately 1.5% better | Improved |
Closest papers by meaning
- Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression (AAAI 2020 · similarity 0.87)
- Bounding Box Regression With Uncertainty for Accurate Object Detection (CVPR 2019 · similarity 0.87)
- Learning Non-Maximum Suppression (CVPR 2017 · similarity 0.87)
- Accelerating Non-Maximum Suppression: A Graph Theory Perspective (NeurIPS 2024 · similarity 0.86)
- Union-over-Intersections: Object Detection beyond Winner-Takes-All (ICLR 2025 · similarity 0.86)
- Bounding Box Stability against Feature Dropout Reflects Detector Generalization across Environments (ICLR 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.
