DALi: Domain Adaptation in LiDAR Point Clouds for 3D Obstacle Detection

The use of modern LiDAR devices for onboard 3D object detection has paved the road for vehicles' full automation. Deep learning approaches leveraging the high-resolution spatial data from laser sweeps have become state-of-the-art in the field, providing an accurate representation of the traffic situation. Nonetheless, the need for a vast amount of annotated samples prevents a smooth deployment of these methods on custom sensor configurations, as differences in the scene geometry or the scanner specifications yield to major performance drops, especially in networks using the raw point cloud as input. To close this gap, an unsupervised domain adaptation training strategy is presented, so that features encoded by PointNet-like backbones are agnostic to the input domain. Through auxiliary domain classifiers and Gradient Reversal Layers, the accuracy of the model on the target domain is significantly boosted without the use of additional labels. Results on the KITTI and nuScenes benchmarks show a notable improvement over the baseline, demonstrating the suitability of GRLs for domain adaptation of 3D detectors based on raw LiDAR point clouds.