Learned Non-Maximum Suppression for 3D Object Detection

2026-06-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

This work introduces two learned filtering modules, D2D-Rescore and GossipNet3D, designed to replace heuristic non-maximum suppression (NMS) in LiDAR-based 3D object detection. D2D-Rescore utilizes transformer-based detection-to-detection (D2D) attention, while GossipNet3D adapts the 2D GossipNet concept for 3D through localized message passing in a bird's-eye view. A key component is a metric-aware matching strategy, aligned with the nuScenes evaluation protocol, which ensures consistent training and validation. Both D2D-Rescore and GossipNet3D significantly improve mean average precision (mAP), nuScenes detection score (NDS), and true positive quality compared to CircleNMS, particularly benefiting small and infrequent object classes. These enhancements are achieved with minimal computational overhead, demonstrating that learned, detection-level filtering can boost 3D detector reliability without altering the base network.

Key takeaway

For Machine Learning Engineers optimizing LiDAR-based 3D object detection, consider integrating learned NMS modules like D2D-Rescore or GossipNet3D. These methods offer a principled alternative to heuristic suppression, significantly boosting mean average precision and nuScenes detection score, particularly for small or infrequent objects. You can enhance detector reliability without modifying your base network, providing a direct path to performance gains in critical perception systems.

Key insights

Learned filtering modules, D2D-Rescore and GossipNet3D, enhance 3D object detection by replacing heuristic NMS with detection-to-detection relation leveraging.

Principles

• Learned filtering improves 3D detection reliability.
• Relational reasoning among detections is effective.
• Metric-aware matching aligns training with evaluation.

Method

D2D-Rescore uses transformer-based D2D attention; GossipNet3D adapts 2D GossipNet to 3D via localized message passing in bird's-eye view.

In practice

• Apply learned NMS post-processing.
• Improve detection for small, infrequent objects.
• Integrate with existing 3D base networks.

Topics

• 3D Object Detection
• Non-Maximum Suppression
• LiDAR Perception
• D2D-Rescore
• GossipNet3D
• nuScenes Benchmark

Code references

• rst-tu-dortmund/learned-3d-nms

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Non-Maximum Suppression using Numpy — Vectorized NumPy implementation of Non-Maximum Suppression efficiently cleans redundant object detection bounding boxes.
• RCGDet3D: Rethinking 4D Radar-Camera Fusion-based 3D Object Detection with Enhanced Radar Feature Encoding — Optimizing 4D radar feature extraction significantly boosts 3D object detection accuracy and speed, outperforming complex fusion methods.
• Learning with Semantic Priors: Stabilizing Point-Supervised Infrared Small Target Detection via Hierarchical Knowledge Distillation — Hierarchical knowledge distillation with VFM semantics stabilizes point-supervised infrared small target detection.
• Revisiting Token Compression for Accelerating ViT-based Sparse Multi-View 3D Object Detectors — Dynamic patch size adjustment and semantic preservation accelerate ViT-based 3D object detection without accuracy loss.
• Boosting Zero-Shot 3D Style Transfer with 2D Pre-trained Priors — Integrating 2D pre-trained decoders into 3D style transfer pipelines effectively mitigates data scarcity and enhances stylization quality.
• A Probabilistic Framework for Improving Dense Object Detection in Underwater Image Data via Annealing-Based Data Augmentation — A pseudo-simulated annealing data augmentation method improves underwater object detection in crowded, natural scenes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.