Unsupervised Deep Learning for Primitive-Based Shape Abstraction

Sammanfattning: The ability to give simple but representative descriptions of complex shapes, known as shape abstraction, is a useful tool in computer science. An unsupervised deep neural network that predicts a set of primitive shapes to describe an arbitrary three-dimensional input shape is constructed and trained on airplane and chair models. The network is based on the research paper by D. Paschalidou et al., with a few modifications to rectify unsatisfactory results. The network is able to predict sets of primitives that resemble the shape of the input. However, the sets of primitives are not parsimonious, i.e. several primitives may be used to describe the same part of an object. A remedy for this problem is proposed through an addition of a punishing function for overlapping primitives, but fail to produce the desired effect. Results from a direct implementation of the network as well as with the modifications are presented and discussed.