Self‐supervised non‐rigid structure from motion with improved training of Wasserstein GANs

This study proposes a self-supervised method to reconstruct 3D limbic structures from 2D landmarks extracted single view. The loss of self-consistency can be reduced by performing random orthogonal projection the reconstructed structure. Thus, training process using geometric in reconstruction–projection–reconstruction process. network mainly consists graph convolution and Transformer encoders. is called SS-Graphformer. By adding discriminator, SS-Graphformer used as generator form Wasserstein Generative Adversarial Network architecture with Gradient Penalty improve accuracy reconstruction. It experimentally demonstrated that addition structure discriminator significantly