Learned Non-Maximum Suppression for 3D Object Detection
2026-06-02 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine LearningRobotics
AI summaryⓘ
The authors propose two new methods to improve the step of filtering overlapping 3D object detections from LiDAR data, which is usually done by a simple rule-based method called non-maximum suppression (NMS). Their methods use learned models to better understand relationships between detected objects, making filtering smarter and more reliable. They show that these learned filters improve detection accuracy, especially for small or rare objects, without slowing down the system much. Importantly, their approach works with existing detectors and matches well with standard evaluation protocols.
LiDAR3D object detectionnon-maximum suppressiontransformerBird's-eye viewmean average precisionnuScenesmessage passingdetection filteringGossipNet
Authors
Timo Osterburg, Stefan Schütte, Torsten Bertram
Abstract
Post-processing is a critical stage in LiDAR-based 3D object detection, where dense and overlapping proposals must be filtered for compact and reliable perception. This work introduces two learned filtering modules that replace heuristic non-maximum suppression (NMS) by leveraging relations among detections. D2D-Rescore employs transformer-based detection-to-detection (D2D) attention, while GossipNet3D adapts the 2D GossipNet concept to 3D through localized message passing in bird's-eye view. A metric-aware matching strategy aligned with the nuScenes evaluation protocol ensures consistent training and validation behavior, improving overall detection performance. Both approaches improve mean average precision (mAP), nuScenes detection score (NDS), and true positive quality compared to CircleNMS, particularly for small and infrequent classes, while adding minimal computational overhead. These results demonstrate that learned, detection-level filtering can enhance 3D detector reliability without modifying the base network, offering a principled alternative to heuristic suppression. Code is available at https://github.com/rst-tu-dortmund/learned-3d-nms .