Scene flow tracks the 3D motion of each point in adjacent point clouds. It provides fundamental 3D motion perception for autonomous driving and server robot. Although red green blue depth (RGBD) camera or light detection and ranging (LiDAR) capture discrete 3D points in space, the objects and motions usually are continuous in the macroworld. That is, the objects keep themselves consistent as they flow from the current frame to the next frame. Based on this insight, the generative adversarial networks (GAN) is utilized to self‐learn 3D scene flow without ground truth. The fake point cloud is synthesized from the predicted scene flow and the point cloud of the first frame. The adversarial training of the generator and discriminator is realized through synthesizing indistinguishable fake point cloud and discriminating the real point cloud and the synthesized fake point cloud. The experiments on Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset show that our method realizes promising results. Just as human, the proposed method can identify the similar local structures of two adjacent frames even without knowing the ground truth scene flow. Then, the local correspondence can be correctly estimated, and further the scene flow is correctly estimated. An interactive preprint version of the article can be found here: https://www.authorea.com/doi/full/10.22541/au.163335790.03073492.

3D scene flow characterizes how the points at the current time flow to the next time in the 3D Euclidean space, which possesses the capacity to infer autonomously the non-rigid motion of all objects in the scene. The previous methods for estimating scene flow from images have limitations, which split the holistic nature of 3D scene flow by estimating optical flow and disparity separately. Learning 3D scene flow from point clouds also faces the difficulties of the gap between synthesized and real data and the sparsity of LiDAR point clouds. In this paper, the generated dense depth map is utilized to obtain explicit 3D coordinates, which achieves direct learning of 3D scene flow from 2D images. The stability of the predicted scene flow is improved by introducing the dense nature of 2D pixels into the 3D space. Outliers in the generated 3D point cloud are removed by statistical methods to weaken the impact of noisy points on the 3D scene flow estimation task. Disparity consistency loss is proposed to achieve more effective unsupervised learning of 3D scene flow. The proposed method of self-supervised learning of 3D scene flow on real-world images is compared with a variety of methods for learning on the synthesized dataset and learning on LiDAR point clouds. The comparisons of multiple scene flow metrics are shown to demonstrate the effectiveness and superiority of introducing pseudo-LiDAR point cloud to scene flow estimation.

No abstract