The work gives attention to learning on human body triangular meshes and is about transferring known deep learning techniques to Non-Euclidean data using spectral mesh processing. Using the Graph Laplacian, meshes geometry is transformed into spectral coefficients that are given to an Autoencoder model. The network is able to give state of the art results by being fast and by allowing generation of new realistic samples.

The code received the Graphics Replicability Stamp and is available with a pretrained model here. Preprint is available here.

Lemeunier, C., Denis, F., Lavoué, G., & Dupont, F. (2022). Representation learning of 3D meshes using an Autoencoder in the spectral domain. In Computers & Graphics (Vol. 107, pp. 131–143). Elsevier BV. https://doi.org/10.1016/j.cag.2022.07.011.