欠損ノイズが再現可能なSim2RealによるLiDARセグメンテーション

抄録

In 3D scene understanding tasks using LiDAR data, constructing training data poses a challenge due to its high annotation cost. To this end, annotation-free simulator-based training has recently been gaining attention, while the domain gap between simulators and real environments often leads to decreased generalization performance. This paper introduces a Sim2Real domain adaptation method that mitigates the domain gap by reproducing realistic raydrop noise onto labeled simulation data using deep generative models, enhancing its applicability to real-world scenarios. We demonstrate the effectiveness of our approach in multiple segmentation tasks.