Core Concepts
LoCondA, a novel framework, generates high-quality 3D meshes from raw point clouds by leveraging a continuous atlas paradigm that preserves local consistency around patch vertices, mitigating the issue of discontinuities in reconstructed shapes.
Abstract
The paper introduces LoCondA, a framework for generating and reconstructing 3D meshes from raw point clouds. It consists of two main components:
Generative auto-encoder using hypernetworks: LoCondA leverages existing solutions like HyperCloud and HyperFlow to map point clouds into a known prior distribution, which can then be sampled to obtain points on the object's surface.
Locally Conditioned Atlas (LoCondA): This component implements the novel "Continuous Atlas" paradigm, which models a local surface structure using a single transformation function ϕ that maps a 2D patch and a point on the object's surface to the local neighborhood of that point. This approach ensures local consistency around patch vertices, addressing the issue of discontinuities in reconstructed shapes observed in previous atlas-based methods.
The key advantages of LoCondA are:
It can generate an arbitrary number of patches to form a watertight mesh, overcoming the limitations of discrete atlas representations.
The local conditioning of the ϕ function ensures smooth stitching of patches, eliminating discontinuities in the final mesh.
Empirical evaluation shows LoCondA produces high-quality, watertight meshes comparable to state-of-the-art methods, while preserving the generative capabilities of the underlying autoencoder.