Program

Alec Jacobson

University of Toronto, Canada

Title: Geometry Processing in The Wild

Abstract: Geometric data abounds, but our algorithms for geometry processing are failing. Whether from medical imagery, free-form architecture, self-driving cars, or 3D-printed parts, geometric data is often messy, riddled with "defects" that cause algorithms to crash or behave unpredictably. The traditional philosophy assumes geometry is given with 100% certainty and that algorithms can use whatever discretization is most convenient. As a result, geometric pipelines are leaky patchworks requiring esoteric training and involving many different people.

Instead, we adapt fundamental mathematics to work directly on messy geometric data. As an archetypical example, I will discuss how to generalize the classic formula for determining the inside from the outside of a curve to messy representations of a 3D surface. This work helps keep the geometry processing pipeline flowing, as validated on our large-scale geometry benchmarks.

Our long term vision is to replace the current leaky geometry processing pipeline with a robust workflow where processing operates directly on real geometric data found "in the wild". To do this, we need to rethink how algorithms should gracefully degrade when confronted with imprecision and uncertainty. Our most recent work on differentiable rendering and geometry-based adversarial attacks on image classification demonstrates the potential power of this philosophy.

Bio: Alec Jacobson is an Assistant Professor and Canada Research Chair in the Departments of Computer Science and Mathematics at University of Toronto. Before that he was a post-doctoral researcher at Columbia University working with Prof. Eitan Grinspun. He received a PhD in Computer Science from ETH Zurich advised by Prof. Olga Sorkine-Hornung, and an MA and BA in Computer Science and Mathematics from the Courant Institute of Mathematical Sciences, New York University. His PhD thesis on real-time deformation techniques for 2D and 3D shapes was awarded the ETH Medal and the Eurographics Best PhD award. Leveraging ideas from differential geometry and finite-element analysis, his work in geometry processing improves exposure of geometric quantities, while his novel user interfaces reduce human effort and increase exploration. He has published several papers in the proceedings of SIGGRAPH. He leads development of the widely used geometry processing library, libigl, winner of the 2015 SGP software award. In 2017, he received the Eurographics Young Researcher Award

Talk slides: Link

Leonidas Guibas

Stanford University, USA

Title: Exploiting Compositional Structure for Shape Generation

Abstract: TBA

Bio: Professor Guibas heads the Geometric Computation group in the Computer Science Department of Stanford University. He is acting director of the Artificial Intelligence Laboratory and member of the Computer Graphics Laboratory, the Institute for Computational and Mathematical Engineering (iCME) and the Bio-X program. His research centers on algorithms for sensing, modeling, reasoning, rendering, and acting on the physical world. Professor Guibas' interests span computational geometry, geometric modeling, computer graphics, computer vision, sensor networks, robotics, and discrete algorithms --- all areas in which he has published and lectured extensively.

Talk slides: TBA

Jiajun Wu

MIT University, USA

Title: Integrating learning with graphics for 3D shape reconstruction, generation, and manipulation

Abstract: Human perception is beyond recognition and reconstruction. From a single image, we're able to explain what we see, reconstruct the scene in 3D, and also imagine how the image would look like if the objects within are in a different position or texture. In this talk, I will present our recent work on reconstructing, generating, and manipulating objects and scenes from visual input. The core idea is to exploit the generic, causal structure behind the world, especially knowledge from computer graphics, and to integrate it with deep learning. I'll focus on a few topics to demonstrate this idea: reconstructing shapes from a single image for objects outside training categories, generating shapes and their corresponding texture, and integrating reconstruction and generation for 3D-aware scene manipulation.

Bio: Jiajun Wu is a fifth-year PhD student at MIT, advised by Bill Freeman and Josh Tenenbaum. He received his undergraduate degree from Tsinghua University, working with Zhuowen Tu. He has also spent time at research labs of Microsoft, Facebook, and Baidu. His research has been supported by fellowships from Facebook, Nvidia, Samsung, Baidu, and Adobe. He study machine perception, reasoning, and its interaction with the physical world, drawing inspiration from human cognition.

Talk slides: Link

Shuran Song

Princeton University, USA

Title: 3D Scene Understanding with Deep Generative Models

Abstract: Intelligent robots require advanced vision capabilities to perceive and interact with the real physical world. In this talk, I will advocate the use of complete 3D scene representations that enable intelligent systems to not only recognize what is seen (e.g. Am I looking at a chair?), but also predict contextual information about the complete 3D environment beyond visible surfaces (e.g. What could be behind the table? Where should I look to find an exit? Where is the unobserved lighting sources?). As examples, I will present two of our recent works that demonstrate the power of these representations through analyzing and synthesizing 3D scenes outside the image field of view (Im2Pano3D), and estimating the lighting information of an environment (Neural Illumination). Finally, I will discuss some ongoing efforts on how these 3D scene representations can further benefit from real-world robotic interactions, shifting the way we view computer vision problems from the perspective of a passive observer to that of an active explorer.

Bio: Shuran Song will be joining the computer science department at Columbia University as an assistant professor starting Fall 2019. Her research interests lie broadly in artificial intelligence, with emphasis on computer vision and robotics. Her work focuses on establishing the complete research infrastructure for the field of 3D visual scene understanding: from developing fundamental algorithms to deploying them in practical real-world robotics applications; from constructing large-scale 3D datasets to designing effective 3D data representations.

Talk slides: TBA

Ali Eslami

Google DeepMind

Title: Neural scene representation and rendering

Abstract: Scene representation—the process of converting visual sensory data into concise descriptions—is a requirement for intelligent behavior. Recent work has shown that neural networks excel at this task when provided with large, labeled datasets. However, removing the reliance on human labeling remains an important open problem. In this talk I will introduce the Generative Query Network (GQN), a framework within which machines learn to represent scenes using only their own sensors. The GQN takes as input images of a scene taken from different viewpoints, constructs an internal representation, and uses this representation to predict the appearance of that scene from previously unobserved viewpoints. The GQN demonstrates representation learning without human labels or domain knowledge, paving the way toward machines that autonomously learn to understand the world around them.

Bio: Ali Eslami is a staff research scientist at Google DeepMind working on problems related to artificial intelligence. His group's research is focused on figuring out how we can get computers to learn with less supervision. Previously he was a post-doctoral researcher at Microsoft Research Cambridge. Hi did his PhD at the University of Edinburgh, where he was a Carnegie scholar. His supervisor was Christopher Williams. During his PhD he was also a visiting researcher at Oxford University working with Andrew Zisserman. He also did an internship at Microsoft Research working with John Winn. He helped organise the PASCAL Visual Object Classes challenge from 2012 to 2015.

Talk slides: Link

Mathieu Aubry

ENPC

Title: Learning surface generation and matching

Abstract: In this talk I will first discuss how deep networks can be used to learn parametric surface deformations for 3D object matching (3D-CODED, ECCV18) and surface generation (AtlasNet, CVPR18). I will then present more recent work exploring how these idea can be used either without basic template or to perform template discovery.

Bio: Mathieu Aubry is a researcher in Ecole des Ponts ParisTech. He obtained his PhD under the supervision of Josef Sivic (INRIA) and Daniel Cremers (TUM) from ENS in 2015, working on 3D shape representations. He then spent a year in UC Berkeley working with Alexei Efros.

Talk slides: Link

Julien Valentin

Google

Title: Deep Generative Modeling with TensorFlow Graphics

Abstract: Despite tremendous progress in both photo-realistic generative modelling and training interpretable generative models, there is still a gap between the two. Indeed, learning photo-realistic and controllable generative models remains an open, but extremely exciting challenge. Computer graphics is a very mature field which offers a plethora of techniques allowing to generate photo-realistic environments which artists can easily modify. Interestingly, a large number of these tools and functionalities are differentiable, making them very attractive to use in deep architecture. In this talk, I will introduce TensorFlow Graphics, a new library aiming at providing differentiable computer graphics layers to stimulate research at the intersection of machine learning and computer graphics.

Bio: Julien Valentin obtained his PhD from Oxford under the supervision of Phil Torr, with a focus on real-time 3D scene understanding, machine learning, and camera pose optimization. Immediately after finishing his PhD, Julien became a founding member of perceptiveIO where he focused on high-speed depth estimation and generative model fitting. He is now continuing his research as a Research Scientist at Google.

Talk slides: Link