dong, hao (xi'an jiaotong university), Zhang, Haolin (Xi'an Jiaotong University), Chen, Shitao (Xi'an Jiaotong University, Xi'an, China), Xin, Jingmin (Xi'an Jiaotong University), Zheng, Nanning (Xi'an Jiaotong University)

CenterNext: Learning Generalized Gaussian Representation for LiDAR 3D Object Detection

Scheduled for presentation during the Regular Session "S37b-Reliable Perception and Robust Sensing for Intelligent Vehicles" (FR-EA-T37), Friday, November 21, 2025, 13:30−13:50, Coolangata 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Center-based 3D object detectors have become increasingly popular due to their advantageous balance between speed and accuracy. However, by relying on standard 2D Gaussian distributions, these methods overlook the representation of object shape and rotation in their learning objectives. This presents challenges to the stability of object detection in practical applications. To address the issue, this work proposes a more generalized Gaussian representation of objects, upon which a novel center-based 3D object detection framework named CenterNext is introduced. Specifically, the proposed method designs a generalized 2D Gaussian kernel to construct object heatmaps. By preserving the unambiguous characteristics of the invariant object center and incorporating variations in object shape and rotation into the formulation, the biases introduced by the previous insufficient object representation are mitigated. The proposed method effectively guides models to recognize the distinctiveness of objects during optimization, thereby enhancing the generalization performance of 3D object detection. Extensive experiments conducted on KITTI, Waymo, and ONCE demonstrate the effectiveness of our approach. With the newly proposed CenterNext, the existing center-based detectors are elevated to a higher level.