Figure 3:
Influence of the sampling density on a detail (top car) of the Kitti frames of Fig. 1.b. First row: On sub-sampled 3D scenes, the regularizing priors of the D-RobOT architecture prevent over-fitting to the random sampling patterns of the target point cloud (blue). The D-RobOT output (last column, red) is very close to the ground truth scene flow (green). Second row: Increasing the number of points per frame reduces the influence of sampling artifacts. The simple RobOT baseline (fourth column, red) still over-fits to the target (blue) but becomes remarkably competitive.