Abstract
Any solid object can be decomposed into a collection of convex polytopes (inshort, convexes). When a small number of convexes are used, such adecomposition can be thought of as a piece-wise approximation of the geometry.This decomposition is fundamental to real-time physics simulation in computergraphics, where it creates a unifying representation of dynamic geometry forcollision detection. A convex object also has the property of beingsimultaneously an explicit and implicit representation: one can interpret itexplicitly as a mesh derived by computing the vertices of a convex hull, orimplicitly as the collection of half-space constraints or support functions.Their implicit representation makes them particularly well suited for neuralnetwork training, as they abstract away from the topology of the geometry theyneed to represent. We introduce a network architecture to represent a lowdimensional family of convexes. This family is automatically derived via anautoencoding process. We investigate the applications of the network includingautomatic convex decomposition, image to 3D reconstruction, and part-basedshape retrieval.