ePointDA: An End-to-End Simulation-to-Real Domain Adaptation Framework for LiDAR Point Cloud Segmentation

Due to its robust and precise distance measurements, LiDAR plays an important role in scene understanding for autonomous driving. Training deep neural networks (DNNs) on data requires large-scale point-wise annotations, which are time-consuming expensive obtain. Instead, simulation-to-real domain adaptation (SRDA) trains a DNN using unlimited synthetic with automatically generated labels transfers the learned model real scenarios. Existing SRDA methods point cloud segmentation mainly employ multi-stage pipeline focus feature-level alignment. They require prior knowledge of real-world statistics ignore pixel-level dropout noise gap spatial feature between different domains. In this paper, we propose novel end-to-end framework, named ePointDA, address above issues. Specifically, ePointDA consists three modules: self-supervised rendering, statistics-invariant spatially-adaptive alignment, transferable learning. The joint optimization enables bridge shift at by explicitly rendering spatially aligning features domains, without requiring statistics. Extensive experiments adapting from GTA-LiDAR KITTI SemanticKITTI demonstrate superiority segmentation.