Cluster-NMS: Improving Crowded Object Detection through Clustering Pattern

• Published in: Signal, image and video processing Vol. 19; no. 9
• Main Authors: Liao, Junguo, Tian, Haonan
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Algorithms; Boxes; Clustering; Computer Imaging; Computer Science; Datasets; Image Processing and Computer Vision; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Signal,Image and Speech Processing; Vision; Anchor-based object detectors; Crowded scenes; Non-maximum suppression; Object detection
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-025-04374-3

Crowded object detection is a highly challenging task in the field of object detection. One major issue is that the standard NMS algorithm encounters difficulties in crowded scenes, as it often erroneously suppresses highly overlapped predictions. In fact, if there are highly clustered bounding boxes in a region, they usually correspond to a ground-truth object. Based on this pattern, we propose an effective cluster-NMS algorithm which can enhance the confidence of not being suppressed for a highly scored bounding box by taking into account the surrounding clustered boxes. It ensures that true predictions are correctly retained in the non-maximum suppression step. On the CrowdHuman dataset, cluster-NMS outperforms other NMS algorithms in improving the performance of crowded object detection. The YOLOv7 detector with cluster-NMS achieves 91.4% AP, which represents the best performance among the crowded object detectors evaluated in the comparative experiments. Furthermore, reliable improvements on CityPersons datasets demonstrate the robustness of our method.