DeepImagetry: Deep Learning 3D Geometry Using 2D Images

Currently, the vision community focuses on extrinsic representation of 3D shapes suitable for learning rigid shapes, whereas the geometry community focuses on adapting convolutional neural networks (CNNs) to non-Euclidean manifolds using intrinsic shape properties for creating optimal descriptors. A method to unify these two complementary approaches has remained elusive. CNN architectures operate on planar structures, which is a challenge for 3D model objects; however, connectivity between different surfaces is not preserved, which is vital for holistic shape analysis.

Researchers at Purdue University have developed the means of transferring 2D images and recreating 3D forms. By learning shapes of 3D images, the developed program can classify and recreate these images for 3D modelling used for differential geometry or shape analysis. This shape recovery is used with the aid of CNN architecture, which allows for fast computing, efficient learning, and sharp mesh creation, allowing for an easier transition from design to 3D form for efficient forensic science and greater special analysis for virtual reality technology.

Advantages

-Generates 3D shape from a single 2D image

-Simplifies 3D tasks such as noise removal or mesh morphing in the 2D domain

Potential Applications:

-Facial recognition

-3D modelling

-Forensics

TRL: 5

Intellectual Property:

Keywords: 3D shape generation from 2D image, CNN architecture for 3D modeling, 2D to 3D reconstruction, deep learning 3D geometry, non-Euclidean manifolds CNNs, shape recovery, forensic science 3D modeling, virtual reality technology analysis, mesh morphing 2D domain, facial recognition 3D, 3D Imaging, 3D Modeling, Computer Technology, Imaging