Overview of the proposed unsupervised adversarial learning framework of 3D Scene Flow. The point clouds of consecutive frames (purple point cloud \(PC_1\) and green point cloud \(PC_2\)) are fed to the scene flow generator \(G_{sf}\), and the output is the 3D scene flow SF for each point in point cloud \(PC_1\), with \(\theta\) being the learnable parameter of \(G_{sf}\). The predicted point cloud \(PC_2^*\) of the second frame is generated by scene flow warping (\(PC_1\) + \(SF\)). Generator loss \(\mathscr{L}_{G}\) and discriminator loss \(\mathscr{L}_{D}\) are designed through the probabilities obtained from the point cloud discriminator, which are used to optimize the scene flow generator and point cloud discriminator, respectively.