Neural Shape Parsers for Constructive Solid Geometry.

Area of Science:

• Computer Vision
• Geometric Modeling
• Deep Learning

Background:

• Constructive Solid Geometry (CSG) defines complex shapes using boolean operations on primitives.
• Parsing shapes into CSG programs offers compact and interpretable generative models.
• The vast space of primitives and combinations presents a significant challenge.

Purpose of the Study:

• Introduce CSGNet, a deep network architecture for generating CSG programs from 2D/3D shapes.
• Address the challenges of parsing shapes into CSG representations.
• Improve upon existing methods for shape modeling and primitive detection.

Main Methods:

• CSGNet employs a convolutional encoder and recurrent decoder for feed-forward shape-to-program mapping.
• Investigated two architectures: CNN-RNN and an augmented version with an explicit memory stack.
• Utilized policy gradient techniques for training on datasets without program annotations.

Main Results:

• CSGNet significantly outperforms bottom-up approaches in speed.
• Stack augmentation improved reconstruction quality and learning efficiency.
• CSGNet demonstrated superior performance as a shape primitive detector compared to state-of-the-art methods.

Conclusions:

• CSGNet provides an efficient and effective method for generating CSG programs from shapes.
• The stack-augmented architecture enhances model performance and learning.
• CSGNet offers a flexible approach, trainable on unannotated data.