Fig. 4From: LiDAR DNN based self-attitude estimation with learning landscape regularitiesPreprocessing scheme. Each 3D point cloud is transformed to a 2D depth image. The depth image is randomly flipped and slid. Flipping, rotation, and L2 normalization are also applied to the gravity vector. In this example, the virtual yaw variant is \(\Delta \psi\) = 90 degBack to article page