Unsupervised Domain Adaptation for Self-Driving from Past Traversal Features

The article presents a method to improve the accuracy of 3D object detection systems for self-driving cars across diverse driving environments. The approach utilizes repeated traversals of multiple locations and LiDAR scans to guide the adaptation process. The method introduces a lightweight regression head to leverage spatial quantized historical features and stabilize training. The framework shows significant improvements in real-world datasets, especially in detecting pedestrians and distant objects.

Publication date: 22 Sep 2023
Project Page: https://github.com/zhangtravis/Hist-DA
Paper: https://arxiv.org/pdf/2309.12140