Zoom link: https://umass-amherst.zoom.us/j/96678534727
Abstract
In recent years, Machine Learning techniques have revolutionized solutions to longstanding image-based problems, like image classification, generation, semantic segmentation, object detection and many others. However, if we want to be able to build agents that can successfully interact with the real world, those techniques need to be capable of reasoning about the world as it truly is: a tridimensional space. There are two main challenges while handling 3D information in machine learning models. First, it is not clear what is the best 3D representation. For images, convolutional neural networks (CNNs) operating on raster images yield the best results in virtually all image-based benchmarks. For 3D data, the best combination of model and representation is still an open question. Second, 3D data is not available on the same scale as images - taking pictures is a common procedure in our daily lives, whereas capturing 3D content is an activity usually restricted to specialized professionals. This thesis is focused on addressing both of these issues. Which model and representation should we use for generating and recognizing 3D data? What are efficient ways of learning 3D representations from a few examples? Is it possible to leverage image data to build models capable of reasoning about the world in 3D?
Our research findings show that it is possible to build models that efficiently generate 3D shapes as irregularly structured representations. Those models require significantly less memory while generating higher quality shapes than the ones based on voxels and multi-view representations. We start by developing techniques to generate shapes represented as point clouds. This class of models leads to high quality reconstructions and better unsupervised feature learning. However, since point clouds are not amenable to editing and human manipulation, we also present models capable of generating shapes as sets of shape handles -- simpler primitives that summarize complex 3D shapes and were specifically designed for high-level tasks and user interaction. Despite their effectiveness, those approaches require some form of 3D supervision, which is scarce. We present multiple alternatives to this problem. First, we investigate how approximate convex decomposition techniques can be used as self-supervision to improve recognition models when only a limited number of labels are available. Second, we study how neural network architectures induce shape priors that can be used in multiple reconstruction tasks -- using both volumetric and manifold representations. In this regime, reconstruction is performed from a single example -- either a sparse point cloud or multiple silhouettes. Finally, we demonstrate how to train generative models of 3D shapes without using any 3D supervision by combining differentiable rendering techniques and Generative Adversarial Networks.
Co-chairs: Rui Wang, Subhransu Maji.
Members: Evangelos Kalogerakis, Duygu Ceylan.