xMUDA: Cross-Modal Unsupervised Domain Adaptation for 3D Semantic Segmentation

Figure 1: Advantages of cross-modal UDA (xMUDA) in presence of domain gap (day-to-night). On this 3D semantic segmentation example, the UDA Baseline [16] prediction from 2D camera image does not detect the car on the right due to the day/night domain shift. With xMUDA, 2D learns the appearance of cars in the dark from information exchange with the 3D LiDAR point cloud, and 3D learns to reduce false predictions.