Adaptive Confluence: Density-Aware Extension of Non-IoU NMS

Non-maximum suppression (NMS) is a critical post-processing step in object detection, traditionally relying on Intersection over Union (IoU) thresholds to remove duplicate detections. However, IoU-based NMS is sensitive to threshold selection and computationally expensive when large numbers of proposals are produced. Confluence was introduced as a non-IoU alternative that selects bounding boxes based on proximity and confidence weighting, but it has been observed to retain false positives in sparse detection regions. This paper proposes Adaptive Confluence, a local density-aware extension of Confluence that incorporates bounding-box local density into the weighted proximity formulation. By modulating suppression strength according to local bounding-box density, Adaptive Confluence improves false-positive suppression in sparse scenes while preserving detections in crowded or occluded regions. Experiments conducted on RetinaNet using the MS COCO 2017 validation dataset demonstrate improved robustness to sparse detection scenarios, correcting a false-positive failure mode of Confluence across several precision- and recall-oriented metrics, including AP@0.75, APsmall, and ARlarge, while maintaining competitive performance in dense scenes. In addition, qualitative results show improved bounding-box selection and reduced false positives. These findings indicate that incorporating density awareness enhances the robustness of non-IoU suppression methods in diverse detection scenarios.