Variational 3D Mesh Generation of Man-Made Objects

Data-driven 3D shape analysis, reconstruction, and generation is an active research topic that finds many useful applications in the fields of computer games, computer graphics, and augmented/virtual reality. Many of the previous mesh-based generative approaches work on natural shapes such as human faces and bodies and little work targets man-made objects. This work proposes a generative probabilistic framework for 3D man-made mesh shapes. Specifically, it proposes a Variational Autoencoder that works directly on mesh vertices and encodes meshes into a probabilistic, smooth, and traversable latent space that can be sampled after training and decoded to generate novel and plausible shapes. Extensive experiments show the representational power of the proposed framework and the underlying latent space. Operations such as random sample generation, linear interpolation, and shape arithmetic can be performed using the proposed method and produce plausible results. An additional advantage of the proposed framework is that it learns to produce a disentangled shape representation which gives finer control over the generated mesh and allows generating shapes with specific qualities without losing the reconstruction power of the autoencoder.