Adversarial unsupervised domain adaptation for 3D semantic segmentation with multi-modal learning

摘要

Semantic segmentation in 3D point-clouds plays an essential role in various applications, such as autonomous driving, robot control, and mapping. In general, a segmentation model trained on one source domain suffers a severe decline in performance when applied to a different target domain due to the cross-domain discrepancy. Various Unsupervised Domain Adaptation (UDA) approaches have been proposed to tackle this issue. However, most are only for uni-modal data and do not explore how to learn from the multi-modality data containing 2D images and 3D point clouds. We propose an Adversarial Unsupervised Domain Adaptation (AUDA) based 3D semantic segmentation framework for achieving this goal. The proposed AUDA can leverage the complementary information between 2D images and 3D point clouds by cross-modal learning and adversarial learning. On the other hand, there is a highly imbalanced data distribution in real scenarios. We further develop a simple and effective threshold-moving technique during the final inference stage to mitigate this issue. Finally, we conduct experiments on three unsupervised domain adaptation scenarios, ie., Country-to-Country (USA.Singapore), Day-to-Night, and Dataset-to-Dataset (A2D2 - SemanticKITTI). The experimental results demonstrate the effectiveness of proposed method that can significantly improve segmentation performance for rare classes. Code and trained models are available at https://github.com/weiliu-ai/auda.