Representing point clouds with generative conditional invertible flow networks

In this paper, we propose a simple yet effective method to represent point clouds as sets of samples drawn from cloud-specific probability distribution. This interpretation matches intrinsic characteristics clouds: the number points and their ordering within cloud is not important all are proximity object boundary. We postulate each parameterized distribution in space, which defined by generative neural network. The network operates composing several spatial transformations locations. Once trained, it provides natural framework for manipulation. For instance can decouple shape its orientation provide routines aligning new into default orientation. To exploit similarities between same-class objects improve model performance, turn weight sharing: networks that densities belonging same family share parameters with exception small, object-specific embedding vector. show these vectors capture semantic relationships objects. Our leverages invertible flow learn embeddings well generate clouds. Thanks formulation contrary similar approaches, able train our an end-to-end fashion. As result, offers competitive or superior quantitative results on benchmark datasets, while enabling unprecedented capabilities perform manipulation tasks, such registration regeneration,