Title: Learned Non-Maximum Suppression for 3D Object Detection

Abstract:

In LiDAR-based 3D object detection, post-processing plays a pivotal role in filtering dense and overlapping proposals to ensure perception is both compact and reliable. This study presents two novel learned filtering modules designed to supplant traditional heuristic non-maximum suppression (NMS) by utilizing the relationships inherent among detections. The first approach, D2D-Rescore, utilizes transformer-based detection-to-detection (D2D) attention mechanisms. The second, GossipNet3D, translates the 2D GossipNet framework into a 3D context via localized message passing within a bird’s-eye view representation. To maintain alignment with the nuScenes evaluation protocol, we implement a metric-aware matching strategy that ensures consistency between training and validation phases, thereby boosting overall detection efficacy.

Compared to CircleNMS, both proposed methods yield enhancements in mean average precision (mAP), nuScenes detection score (NDS), and true positive quality, with particularly notable gains for small and infrequent object classes. These improvements are achieved with negligible additional computational cost. The findings highlight that detection-level filtering can significantly bolster the reliability of 3D detectors without necessitating changes to the underlying network architecture, providing a robust, principled substitute for heuristic suppression. The source code is publicly accessible at https://github.com/rst-tu-dortmund/learned-3d-nms.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC