Shape Program

Shape programs represent 3D objects as structured, interpretable code, offering advantages over traditional methods like CAD model retrieval by enabling semantic reasoning and easier editing. Current research focuses on developing differentiable shape program modules, often integrated with neural networks and techniques like Monte Carlo Tree Search, to improve 3D reconstruction from images and point clouds, and to facilitate learning from both supervised and unsupervised data. These advancements are improving the efficiency and accuracy of 3D modeling and manipulation, particularly through the integration of text-based instructions for editing, and hold significant potential for applications in computer graphics, robotics, and AI-driven design.