PCA Retargeting: Encoding Linear Shape Models as Convolutional Mesh Autoencoders
NeurIPS 2020 Workshop on Pre-registration in Machine Learning, PMLR 148:87-99, 2021.
3D Morphable Models have long played a key role in the construction of statistical shape models. While earlier models employed Principal Component Analysis, recent work has migrated towards mesh autoencoder models for the construction of lightweight, non-linear shape models that facilitate state-of-the-art reconstruction and the capture of high-fidelity details. Doing so results in a loss of interpretability and regularisation in the model latent space. To address this, we propose PCA retargeting, a method for expressing linear PCA models as mesh autoencoders and thereby retaining the gaussianity of the latent space. To encourage the capture of mesh details outside the expressive range of a PCA model, we introduce “free” latent space parameters. Experiments demonstrate the successful retargeting of the PCA models as mesh autoencoders. The introduction of “free” latent parameters have a greater impact when smaller latent vector sizes are used, but do not lead to any gains in reconstruction fidelity.