Adaptive Multi-Prototype Grouping Alignment for Domain Adaptive 3D Detection

Shijie Chen, Junbao Zhuo, Bochao Zou, Jiansheng Chen, Huimin Ma

Keywords: domain adaptation, 3d object detection

Abstract: 3D object detection is crucial for autonomous driving and robotics, but models often perform poorly when deployed in new environments due to domain shifts. While 3D unsupervised domain adaptation methods aim to address this issue, they still struggle with two key challenges: insufficient cross-domain feature alignment and ambiguous foreground-background decision boundaries. In this paper, we propose an Adaptive Multi-Prototype Grouping Alignment framework that addresses these challenges. Our method automatically discovers and dynamically updates feature groups, enabling adaptively cross-domain feature alignment in a fine grained manner. Additionally, we develop two techniques to address the issue of ambiguous foreground-background decision boundaries: Noise Background Hybrid Augmentation that leverages labeled source instances to enhance adaptation in uncertain regions, and Noise Foreground Contrastive Learning that improves foreground-background discrimination by pushing low-confidence features away from prototypes. We conduct extensive experiments on multiple cross-domain 3D detection benchmarks and the results demonstrate the superiority of our method over the state-of-the-art methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6466