CSG-Stump: A Learning Friendly CSG-Like Representation for Interpretable Shape Parsing

TL;DR

This paper introduces CSG-Stump Net, an unsupervised neural network that learns interpretable, compact 3D shape representations from point clouds using a novel three-level CSG-like structure, improving shape generation and interpretability.

Contribution

The paper proposes CSG-Stump Net, a novel deep learning architecture that simplifies CSG representations into a regular, interpretable form suitable for neural networks, enhancing shape parsing performance.

Findings

Outperforms previous CSG-based methods in shape generation quality.

Produces more appealing and interpretable 3D shape reconstructions.

Demonstrates the effectiveness of the CSG-Stump structure through extensive experiments.

Abstract

Generating an interpretable and compact representation of 3D shapes from point clouds is an important and challenging problem. This paper presents CSG-Stump Net, an unsupervised end-to-end network for learning shapes from point clouds and discovering the underlying constituent modeling primitives and operations as well. At the core is a three-level structure called {\em CSG-Stump}, consisting of a complement layer at the bottom, an intersection layer in the middle, and a union layer at the top. CSG-Stump is proven to be equivalent to CSG in terms of representation, therefore inheriting the interpretable, compact and editable nature of CSG while freeing from CSG's complex tree structures. Particularly, the CSG-Stump has a simple and regular structure, allowing neural networks to give outputs of a constant dimensionality, which makes itself deep-learning friendly. Due to these characteristics of CSG-Stump, CSG-Stump Net achieves superior results compared to previous CSG-based methods and generates much more appealing shapes, as confirmed by extensive experiments. Project page: https://kimren227.github.io/projects/CSGStump/