Jan. 19, 2021, 2:04 p.m.
Michał Stypułkowski has won BNY Mellon Master's Thesis Competition
Generative models for point clouds are believed to be a promising step towards intelligent systems able to generalize learned information about 3D objects. They are inspired by the human ability to perceive surroundings in terms of class-specific shape distributions.
“Representing Point Clouds with Generative Conditional Invertible Flow Networks” introduces a simple yet effective method to represent point clouds as sets of samples drawn from a cloud-specific probability distribution. More specifically, each cloud is represented as a parameterized probability distribution defined by a generative neural network. To exploit similarities between same-class objects and to improve model performance, weights are shared between the networks that model densities of points belonging to objects in the same family with the exception of a small, object-specific embedding vector.
As a result, the model offers competitive or superior quantitative results on benchmark datasets, while enabling unprecedented capabilities to perform cloud manipulation tasks, such as point cloud reconstruction and alignment.