CVPR Daily - Tuesday

DAILY T u e s d a y Poster Presentation 16 Learning Unsupervised Hierarchical Part Decomposition of 3D Objects From a Single RGB Image Despoina Paschalidou is a fourth year PhD student at the Max Planck Institute for Intelligent Systems in Tübingen and ETH Zürich. She returns to CVPR for a third successive year with a paper that pushes forward the progress already made in modelling the local 3D geometry of an object. She speaks to us ahead of her poster sessions today. Recent advances in deep learning have led to improved techniques for efficiently capturing the local 3D geometry of an object, including methods that learn meshes, pointclouds, voxel grids, or implicit shape representations. Despoina’s paper takes this one step further. It explores the modelling of high-level properties , including relationships between parts of an object, proximity, and symmetry. She believes this is the first step towards representing more complex concepts, such as regularity . The work is focused on primitive-based representations , which represent the geometry of an object as a set of atomic elements called primitives . It is based on the assumption that you use more primitives to capture the details of the more complex parts of an object, and fewer components to capture parts of an object that are less geometrically complex. The key difference between primitive- based representations and other representations that yield voxel grids, pointclouds and meshes has to do with the fact that these representations