Unsupervised Deep Shape from Template

This paper presents Unsupervised Deep Shape from Template (UDSfT), a novel method that leverages deep neural networks (DNNs) for reconstructing the 3D surface of an object using a single image. More specifically, the reconstruction of isometric deformable objects is achieved in the proposed UDSfT method via a DNN-based template-based framework. Unlike previous approaches that leverage supervised learning, the proposed UDSfT method leverages the notion of unsupervised learning to overcome this obstacle and provide real-time 3D reconstruction. More specifically, UDSfT achieves this via an unsupervised structure that leverages a combination of real-data and synthetic data. Experimental results show that the proposed UDSfT method outperforms the state-of-the-art Shape from Template methods in object 3D reconstruction.