SweepNet: unsupervised learning shape abstraction via neural sweepers

Shape abstraction simplifies complex geometric structures while retaining essential features, and sweep surfaces effectively capture object geometry for this purpose. In this thesis, we introduce SweepNet, a novel unsupervised framework for 3D shape abstraction using sweep surfaces. SweepNet first integrates sweep surfaces into shape abstraction, overcoming challenges related to differentiability and parametrization. We propose an efficient parametrization for sweep surfaces, using superellipses for profile representation and B-spline curves for the axis. This compact representation, requiring as few as 14 floating-point numbers, allows for intuitive and interactive editing while preserving shape details. Extensive experiments highlight the superiority of our model and sweep surface primitives in representing curvyfeaturing objects, showcasing its effectiveness in 3D shape abstraction tasks.