In order to make the point cloud discriminator in adversarial learning correctly discriminate whether the input point cloud comes from the generated data or the real data, we propose four kinds of discriminators, as shown in Figure 4. As shown in Table 4, we perform ablation experiments for different point cloud discriminators. The best performance in scene flow estimation is achieved when PC1 and PC2* perform flow embedding and the weights of the set conv layer are shared. This setup is more beneficial to improve the performance of scene flow estimation.