CVPR Daily - Tuesday

DAILY T u e s d a y Workshop 26 Encoding a shape into a latent space, and then reconstructing the shape and its structure. (from Paul Guerrero’s talk, describing StructureNet Mo et al’s work at SIGGRAPH Asia 2019, they have follow up work at CVPR 2020 about editing shapes called StructEdit) At the end of the workshop an exciting panel discussion brought together invited speakers and organizers to address questions from the audience. One of the main themes covered in the discussion was the contrast between “real” 3D data and synthetic 3D data and the degree to which supervised vs unsupervised methods should be a focus for future work. The panelists agreed that focusing on methods towards the unsupervised end of the spectrum and dealing with real data at scale is an important direction for future work. There was consensus that we should rely more on observation from the real world, especially for how humans interact with the world , and progressively rely less on synthetic datasets. At the same time, synthetic datasets allow us to explore exciting directions such as training AI agents in simulation. Some first steps in this direction are taken in the SAPIEN project (to be presented at the main conference) which added mobility information on top of datasets such as PartNet to create interactive objects which can be used to train robots to open and close cabinets. Understanding the 3D structure of the world in this way can enable the use of these generative models to create content for training AI agents as well. Another topic revolved around handling “ rare events ” that are important but rarely observed (e.g., car crashes, heart conditions). Such events are rare in the real world and hard to produce synthetically. Enumerating such “corner case” events is hard. The panelists discussed the distinction between slow , logic-