SweepNet: Unsupervised Learning Shape Abstraction via Neural Sweepers

TL;DR

This work tackles unsupervised 3D shape abstraction by learning compact sweep-surface representations. It introduces a parameterization that uses a $2$D superellipse profile, a $3$D B-spline axis, and a polynomial scaling function, enabling descriptive shapes with as few as 14 floats and facilitating editing. A differentiable neural sweeper estimates occupancy fields $O_i(t)$ for sweep surfaces within an encoder–decoder framework, trained via a reconstruction loss $\ abla\mathcal{L}_{recon}$ defined with a Boltzmann operator and sharpness parameter $\\alpha$, plus regularization terms for overlap, parsimony, and axis alignment. Experiments on GC-Object and quadruped datasets show SweepNet achieves strong quantitative performance and qualitative expressiveness for curvy and tubular geometries, with editing capabilities demonstrated through parameter manipulation. Limitations include difficulties with highly porous or CAD-like shapes, motivating future work to integrate sweep primitives with other primitives (e.g., CSG) and to develop more generalizable models beyond single-shape training.

Abstract

Shape abstraction is an important task for simplifying complex geometric structures while retaining essential features. Sweep surfaces, commonly found in human-made objects, aid in this process by effectively capturing and representing object geometry, thereby facilitating abstraction. In this paper, we introduce \papername, a novel approach to shape abstraction through sweep surfaces. We propose an effective parameterization for sweep surfaces, utilizing superellipses for profile representation and B-spline curves for the axis. This compact representation, requiring as few as 14 float numbers, facilitates intuitive and interactive editing while preserving shape details effectively. Additionally, by introducing a differentiable neural sweeper and an encoder-decoder architecture, we demonstrate the ability to predict sweep surface representations without supervision. We show the superiority of our model through several quantitative and qualitative experiments throughout the paper. Our code is available at https://mingrui-zhao.github.io/SweepNet/

Figures (19)

• Figure 1: Sweep surfaces data samples for neural sweeper training.
• Figure 2: Pipeline overview. The model processes voxel input to extract a skeletal prior and encodes the data with a voxel encoder. The sweep surface head predicts sweep surface parameters: 2D profiles, 3D sweeping axes, and profile scaling function coefficients, conditioned on the skeletal prior. Training involves generating point clouds for each sweep surface through a differentiable sampler, which the neural sweeper uses to estimate their occupancy. This data is then assembled to reconstruct the input shape to quantify loss. At inference time, the sweep surface parameters are directly processed by a non-differentiable, imperative sweeper to produce the resembled shape.
• Figure 2: Additional qualitative results on GC-objects and quadrupeds datasets. Our method better captures curves.
• Figure 3: Sweep surface primitives parameterized with different profiles, axis, and scaling functions. The 2D profiles (superellipses) are shown in blue. Constant and dynamic scales of sweep surfaces are shown in alternating columns with respect to each profile and axis pair.
• Figure 3: Shape abstraction outcomes from SweepNet utilizing sweeping axes defined by 3 and 4 control points, respectively. Primitives characterized by 3 control points are more concise, while those with 4 control points showcase more turning curves and employ fewer primitives in abstraction.