Abstract

3D reconstruction from multiple 2D images is an inherently ill-posed problem. Prior knowledge is required to resolve ambiguities and probabilistic models are desirable to capture the ambiguities in the reconstructed model. In this talk, I will present two recent results tackling these two aspects. First, I will introduce a probabilistic framework for volumetric 3D reconstruction where the reconstruction problem is cast as inference in a Markov random field using ray potentials. Our main contribution is a discrete-continuous inference algorithm which computes marginal distributions of each voxel’s occupancy and appearance. I will show that the proposed algorithm allows for Bayes optimal predictions with respect to a natural reconstruction loss. I will further demonstrate several extensions which integrate non-local CAD priors into the reconstruction process. In the second part of my talk, I will present a novel framework for deep learning with 3D data called OctNet which enables 3D CNNs on high-dimensional inputs. I will demonstrate the utility of the OctNet representation on several 3D tasks including classification, orientation estimation and point cloud labeling. Finally, I will present an extension of OctNet called OctNetFusion which jointly predicts the space partitioning function with the output representation, resulting in an end-to-end trainable model for volumetric depth map fusion.