Reproduce “ParSeNet: A Parametric Surface Fitting Network for 3D Point Clouds” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, Radomír Měch
ParSeNet is an end-to-end trainable deep network developed by researchers at UMass Amherst and Adobe Research that converts 3D point clouds into a collection of editable parametric surface patches, including B-splines. It achieves superior point cloud segmentation and surface reconstruction, demonstrating a 47.64% relative reduction in residual error compared to prior learning-based methods like SPFN.
Key insights
- Incorporating B-spline patches significantly enhances the expressivity and fidelity of 3D point cloud reconstructions, making outputs more compatible with professional CAD modeling practices.
- An end-to-end differentiable pipeline, leveraging mean-shift clustering for flexible segmentation and a specialized network (SplineNet) for B-spline fitting, enables the learning of semantically meaningful shape decompositions.
- A multi-stage training process with tailored loss functions, including permutation-invariant control point regression and Laplacian loss, is crucial for robustly learning complex parametric surface properties like shape and smoothness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.