.png){kind=icon}
.png){kind=icon}
Return
Adaptive spatial-aware non-maximum suppression for dense object detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132541.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Non-Maximum Suppression (NMS) is a crucial post-processing step in object detection, yet its performance often degrades in dense or heavily occluded scenes. Conventional NMS relies solely on the Intersection-over-Union (IoU) metric to suppress redundant detections, which fails to capture spatial discrepancies in scale, position, and shape, leading to the mis-suppression of adjacent true targets. To address these limitations, we propose an Adaptive Spatial-Aware NMS (ASA-NMS). The proposed method introduces a spatial distance-aware similarity measure that replaces IoU with a normalized diagonal distance to model spatial relations between bounding boxes. Furthermore, a confidence-guided dynamic thresholding strategy is designed to adaptively adjust suppression strength, achieving a better balance between redundancy removal and true detection preservation. Without mod ifying the network structure, ASA-NMS can be seamlessly integrated into existing detection pipelines. Extensive experiments on the CrowdHuman, COCO 2017, and Pascal VOC datasets demonstrate that our method signif icantly improves detection precision and recall, particularly in crowded and occluded scenarios. The code is available at https://github.com/JOU-UIP/ASA-NMS
Keywords:
Dense object detection
Non-maxima suppression
Spatial perception
Computer vision
Deep learning