Shape Abstraction is a fundamental area in computer vision to simplify shapes into a set of primitives, enabling structured understanding of a world. While the choice of primitive determines how the world is perceived and interpreted, achieving both h...
Shape Abstraction is a fundamental area in computer vision to simplify shapes into a set of primitives, enabling structured understanding of a world. While the choice of primitive determines how the world is perceived and interpreted, achieving both high quality and versatility of abstraction remains as a challenging problem. In this paper, we propose a novel shape abstraction framework based on differentiable support function (DSF). DSF can represent a wide range of convex shapes using fewer parameters, providing differentiable contact features (gap, point, and normal) that is essential in contact-related applications. To address the optimization and combinatorial complexity for DSF-based shape abstraction, we introduce two main techniques: 1) differentiable shape parameterization to improve abstraction accuracy, and 2) hyperplane-based marching to enhance abstraction efficiency by reducing the number of required DSFs. Through experiments, we demonstrate that our method outperforms existing approaches in terms of accuracy and efficiency, and we further validate the versatility of our method in various contact-related downstream tasks.