2019/08/20 by Roman Klokov, Jakob Verbeek, Klokov, Roman +3
Computer Science · Engineering · Mathematics · #3D Shape Modeling and Analysis #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Medical Image Segmentation Techniques #Morphological variations and asymmetry
paper · pdf · doi:10.48550/arxiv.1908.07475
openalex publication_date 2019/08/20 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We study end-to-end learning strategies for 3D shape inference from images,\nin particular from a single image. Several approaches in this direction have\nbeen investigated that explore different shape representations and suitable\nlearning architectures. We focus instead on the underlying probabilistic\nmechanisms involved and contribute a more principled probabilistic\ninference-based reconstruction framework, which we coin Probabilistic\nReconstruction Networks. This framework expresses image conditioned 3D shape\ninference through a family of latent variable models, and naturally decouples\nthe choice of shape representations from the inference itself. Moreover, it\nsuggests different options for the image conditioning and allows training in\ntwo regimes, using either Monte Carlo or variational approximation of the\nmarginal likelihood. Using our Probabilistic Reconstruction Networks we obtain\nsingle image 3D reconstruction results that set a new state of the art on the\nShapeNet dataset in terms of the intersection over union and earth mover's\ndistance evaluation metrics. Interestingly, we obtain these results using a\nbasic voxel grid representation, improving over recent work based on finer\npoint cloud or mesh based representations.\n