3D Shape Variational Autoencoder Latent Disentanglement via Mini-Batch Feature Swapping for Bodies and Faces

Abstract

Figure 1. Schematic Description of the Proposed Method. Left: an arbitrary identity feature is selected and a mini-batch of vertices (Xij) is created by swapping features across different 3D shapes. Colours represent identities. Notice that features from the same identity have the same colour. Centre: a 3D-VAE ({E, G}) encodes Xij in its latent representations zij = (z f ij |z c ij ), which are subsequently decoded into X ij . In this case f corresponds to the nose. Therefore, while z f ij controls the shape of the nose, z c ij controls the shape of the rest of the face. Right: visual representation of all the different mesh features for which we seek to obtain a disentangled latent representation.