DeepSDF AI gets a 3D introduction from Facebook, MIT, and UW

Author at TechGenyz Facebook

DeepSDF characterizes the symbolic distance function of shape through a potential encoding and feedforward decoder network. When deep convolutional networks are used directly in 3D space, their time and space complexity will increase dramatically. And more classic and compact surface representations (such as triangular or quadrilateral meshes) can cause problems in training because we may need Handles an unknown number of vertices and any topology.

These challenges impose limitations on the quality, flexibility, and fidelity of deep learning methods when attempting to process input 3D data or output 3D reasoning for target segmentation and reconstruction using deep learning methods.

The latest research from Facebook Reality Lab demonstrates an efficient, continuous new generation of 3D modeling characterization and methods. The method uses the concept of Signed Distance Function (SDF). The common surface reconstruction technique discretizes SDF into a regular grid for estimating and measuring denoising. This method learns a generation model to generate continuous fields.

The continuous characterization proposed by this study can be intuitively understood as a learned shape classifier whose decision boundary is the surface of the shape itself. The proposed method, like other studies, attempts to map potential space to 3D complex shape distributions, but the main characterization is different. Although implicit surface SDF is well known in the computer vision and graphics community, there has not been a study of a continuous, generalizable 3D generation model that directly learns SDF.

The contributions of this research with Facebook include 3D modeling of generated shapes using continuous implicit surfaces. 3D shape learning method based on probabilistic self-decoder; the application of this method in shape modeling and completion is demonstrated. The model produces high-quality continuous surfaces with complex topologies and obtains current optimal results in quantitative comparisons of shape reconstruction and completion.