TCE Conference 2016 – Program
Research Workshop of the Israel Science Foundation (ISF)
The 6th Annual Henry Taub International TCE Conference
3D Visual Computing:
Graphics, Geometry & Imaging
May 24-25, 2016
Kogan Auditorium, Meyer Building (EE) Technion, Haifa
Tuesday May 24th
Gathering & registration
Moshe Sidi – Senior Executive Vice President –Technion – Israel Institute of Technology
Gadi Singer, Intel – Vice President PEG, General Manager Intel Israel Development Center, GM IP Group Architecture
Session Chair: Alfred M. Bruckstein
The Rubinger Family Visiting Lectureship in TCE
William T. Freeman, MIT & Google, USA
Learning About Things by Hitting Them with a Drumstick
Marc Pollefeys, ETH Zurich, Switzerland
Semantic 3D Reconstruction
Session Chair: Ron Kimmel
Thomas Funkhouser, Princeton University, USA
Finding 3D Surface Correspondences with Shape Analysis
Ayelet Dominitz-Mashiah, Elbit Systems Ltd, Israel
3D on the Fly
Henry Taub Distinguished Visitor
Niloy Mitra, University College London, UK
Computational Design of Functional Objects
Session Chair: Michael Lindenbaum
The Rubinger Family Visiting Lectureship in TCE
Srinivasa Narasimhan, Carnegie Mellon University, USA
The Science and Engineering of Light Transport
Avram Golbert, Rafael, Israel
What from Where In 3D! Learning Semantic Segmentation from 3D Models
Yaron Lipman, Weizmann Institute of Science, Israel
Drawing Graphs in Orbifolds
Session Chair: Yehoshua Y. Zeevi
Alex M. Bronstein, Technion, Israel
Bayesian Estimation of Correspondence Between Deformable Shapes
16:20-16:40 Tali Treibitz, University of Haifa, Israel
3D in the Ocean
Guy Medan, BrightSource, Israel
17:00-17:40 Ira Kemelmacher-Shlizerman, University of Washington, USA
People Modeling from Historical Footage
Wednesday May 25th
Gathering & registration
Session Chair: Gershon Elber
Helmut Pottmann, TU Wien, Austria
Modeling Architectural Freeform
Marc Alexa, TU Berlin, Germany
Representing Frames as Möbius Transformations
10:30-10:45 Coffee break
Session Chair: Yoav Schechner
Kristen Grauman, University of Texas at Austin, USA
Egomotion and Visual Learning
Mark Kliger, Intel, Israel
Taking a "Deeper" Look at 3D Scenes
The Rubinger Family Visiting Lectureship in TCE
Michael Kazhdan, Johns Hopkins University, USA
Signal Processing – from Images to Surfaces
Session Chair: Ayellet Tal
The Rubinger Family Visiting Lectureship in TCE
Eitan Grinspun, Columbia University, USA
The Geometry of Physics Has a Profound Impact on Computation
Frédéric Parienté, NVIDIA
Advances in GPU Architecture for Deep Learning and Scientific Computing
Yair Weiss, The Hebrew University of Jerusalem, Israel
Statistics of RGBD Images
Session Chairs: Mirela Ben-Chen
Amit Batikoff, Applied Materials, Israel
3D in the Nano-Scale Semiconductors World
Leonid Karlinsky & Yochay Tzur, IBM Research, Israel
Fine-Grained Recognition of Thousands of Object Categories with Single-Example Training
Daniel Dikovsky, Stratasys Ltd., Israel
Multi-Material 3D Printing – Opportunities and Challenges
Closing commentsMarc Alexa, TU Berlin, Germany - Representing Frames as Möbius TransformationsLet us say that a frame is given by three ‘sticks’ (of equal lengths) meeting in one common point. We are interested in representing the orientation and the ‘shape’ of the frame. Orientation is the rotation relative to a reference frame; and ‘shape’ is the deformation relative to a reference frame. It turns out that any frame can be turned into any other frame by a Möbius transformation.
This viewpoint reveals that rotations are points on a 3-sphere, the so-called unit quaternions. Unit quaternions are well-known and quite useful as a representation for rotations in space — they are continuous in the variables, minimal in the sense that at least four coordinates are necessary for a continuous representation, and they come with a natural metric that allows us to measure the ‘amount’ of rotation, i.e. the angle.
The viewpoint of Möbius transformations also reveals, and this is the new aspect of this work, that deformations are points on a hyperboloid. So ‘shape’ can be described as a point in hyperbolic space. This is a representation that, just like unit quaternions, is continuous, small, and comes with a natural metric that allows measuring the amount of deformation.Amit Batikoff, Applied Materials, Israel - 3D in the Nano-Scale Semiconductors WorldAbstract – Due to the increasing challenge in shrinking transistors size the semiconductor industry is changing its direction, structures are being built in a 3D manner for allowing better conductivity and lower leakage current. Only Atomic level control over the fin and gate structures, in the FinFET design will allow the device to gain from the this new technology. Traditional metrology tools are mostly based on Scanning Electron Microscope and provide images from top-down perspective, allowing to measure the lateral dimensions of the structures. Innovation is required in the metrology tools to allow imaging, reconstruction and measuring of the 3D shapes with a sub-nanometer precision.
Bio – Amit Batikoff is managing the High Resolution SEM Inspection and Review algorithm R&D at Applied Materials. He is a Signal Processing and Computer Vision expert in the fields of noise estimation, outlier detection, image reconstruction and enhancement. Amit is Applied Material’s leader in the OMEK consortium focusing on 3D metrology algorithms. Amit holds a MSc in EE from TAU.Alex M. Bronstein, Technion, Israel - Bayesian Estimation of Correspondence Between Deformable ShapesAbstract –Finding correspondences between deformable shapes is one of the fundamental problems in computational shape analysis. The explosive growth of 3D content driven by the increasing affordability of depth sensing technologies makes it particularly acute. While a lot of progress have been done in this field in the past few years, this problem is far from being solved. In this talk, I will show a perspective, from which the correspondence problem is viewed as Bayesian estimation of a latent bijection between two shapes from the realizations of a stochastic process. This view leads to simple (almost trivial) estimation algorithms that lend themselves well to efficient computation and, surprisingly, perform very well in practice.Daniel Dikovsky, Stratasys Ltd., Israel - Multi-Material 3D Printing – Opportunities and ChallengesAbstract – Today we witness the great impact of 3D printing technologies on “how we fabricate” and “what we fabricate”, the latter of which has been fundamentally affected by the freedom of form enabled by the additive manufacturing techniques. Now we witness the elimination of another major fabrication constraint as we add the freedom of material enabled by multi-material 3D printing technology. This fabrication method was pioneered by Stratasys PolyJet printing systems utilizing ink-jet heads that simultaneously deposit liquid droplets of different materials and solidify them using UV light. The deposition sequence and the resulting spatial distribution of the droplets are controlled via computer, comprising a system that turns digital voxels into physical blocks, termed Digital Materials. Consequently, a set of resins with different physical properties can yield numerous Digital Materials, each having a unique set of properties, inherited from its parent materials. This talk will present the capabilities and applications of multi-material 3D printing, focusing on R&D of computational challenges and multi-material CAD approaches.
Bio – Dr. Daniel Dikovsky is a materials scientist and R&D Manager at Stratasys Ltd., a manufacturer of 3D printing equipment and materials for creating physical objects directly from digital data. The Israeli branch of Stratasys (formerly Objet Ltd.) utilizes ink-jet technology for printing three-dimensional polymer objects. Daniel’s research focuses on Multi-Material 3D Printing technology enabling the generation of Digital Materials. These materials are created by simultaneous deposition of multiple material components onto the printing tray. Daniel holds a Ph.D. degree in Biomedical Engineering from The Technion – Israel Institute of Technology and a M.Sc. degree in Applied Chemistry from The Hebrew University of Jerusalem, Israel.Ayelet Dominitz-Mashiah, Elbit Systems Ltd, Israel - 3D on the Fly
Abstract –Elbit Aerospace utilizes 3D mapping in a variety of projects. The talk will highlight two of our projects. First is a 3D model generated during flight and used to compensate for degraded visual environment. In this project, a 3D model is obtained by combining a 3D point cloud scanned by a LiDAR, together with images from a camera to obtain a photorealistic model. Second – the 3D model is generated offline from aerial imaging, then distributed and consumed by various customers. Here, the rapidly changing viewpoint of an aircraft mounted camera enables multi view geometry techniques to obtain a 3D mesh and its texture. In this talk I will present a detailed novel approach for generating multiple-view stereo 3D reconstruction while compensating for rolling shutter artifacts.
Bio – Ayelet (Dominitz) Mashiah leads the image processing and computer vision group at Elbit Systems Airborne division. She holds a PhD in 3D mapping from the Department of Electrical Engineering at the Technion, Israel Institute of Technology. Ayelet has more than 10 years experience in supervising and conducting research and development in the fields of computer vision, computer graphics, medical imaging, image analysis and machine learning.William T. Freeman, MIT & Google, USA - Learning About Things by Hitting Them with a DrumstickAbstract – Children explore their visual world, in part, by pushing, scraping, and banging things. Inspired by that behavior, we sought to learn what we could by recording video and audio of a drumstick striking many different objects. Can we use that data to learn how visual appearance relates to material properties? To pose it as a machine learning task, we sought to predict the audio from silent videos of the drumstick strikes. Successful prediction would require some implicit information about the mechanical characteristics of the viewed materials. I'll describe the dataset, the representations for the video and the audio, and our results. In 2-alternative tests of which audio was recorded from a given video, we fooled human observers 40% of the time. We show that our system, trained only from the video and audio, has implicit knowledge of the material labels in silent test videos. PhD thesis work of Andrew Owens, in collaboration with Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson and William T. Freeman. http://arxiv.org/abs/1512.08512, to appear in CVPR 2016
Bio – William T. Freeman is the Thomas and Gerd Perkins Professor of Electrical Engineering and Computer Science at MIT, and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL) there. His current research interests include machine learning applied to computer vision, Bayesian models of visual perception, and computational photography. He received outstanding paper awards at computer vision or machine learning conferences in 1997, 2006, 2009 and 2012, and test-of-time awards for papers from 1990 and 1995. Previous research topics include steerable filters and pyramids, orientation histograms, the generic viewpoint assumption, color constancy, computer vision for computer games, and belief propagation in networks with loops. He is active in the program or organizing committees of computer vision, graphics, and machine learning conferences. He was the program co-chair for ICCV 2005, and for CVPR 2013.Thomas Funkhouser, Princeton University, USA - Finding 3D Surface Correspondences with Shape AnalysisAbstract – Finding correspondences between regions of 3D shapes is valuable in a number of application domains, including molecular biology, archaeology, paleontology, etc. However, detecting geometrically similar surface regions is often not sufficient. Functionally important correspondences often are revealed only through analysis of large-scale structures, symmetries, affordances, and semantics. This talk will describe a few projects where cues of those types are used to detect surface correspondences that would be difficult to find otherwise.
Bio – Thomas Funkhouser is a professor in the Department of Computer Science at Princeton University. His research focuses on topics in 3D shape analysis in computer graphics, computer vision, and geometric modeling. He is the recipient of a Sloan Foundation Fellowship, a NSF Career Award, two SEAS Teaching Awards, and a SIGGRAPH Computer Graphics Achievement Award.Avram Golbert, Rafael, Israel - What from Where In 3D! Learning Semantic Segmentation from 3D ModelsAbstract – We address the challenge of classifying pixels in aerial images of urban areas which can provide crucial data for various applications such as mapping, 3D modeling and navigation. Fully Convolutional Networks are very well suited to handle image segmentation; however generating enough annotated data is very costly. We present a semi-supervised system to annotate an entire urban 3D model generated from Multi- View Stereo. The annotated model is then back projected onto thousands of images that are then used to train the network. The exact world location of every pixel, provided by the model, enables a variety of geometric features, simplifying the semantic annotation. We introduce a novel loss function to optimize the F1 score calculated over the entire batch. We present results on multiple data-sets and compare results with and without depth information.
Bio – Avram Golbert leads Rafael’s 3D Understanding group, which researches 3D scene reconstruction and the applications of geometric understanding to algorithms such as mapping, real time urban navigation and object detection. Avram holds a BSc in Mathematics and MSc in Computer Science from the Hebrew University, where he researched object detection using geometric and semantic context. This work was funded in part by the Omek consortium and was done in part while Avram was a guest researcher at the Deep Vision Lab in Tel Aviv University with Prof. Lior WolfKristen Grauman, University of Texas at Austin, USA - Egomotion and Visual LearningAbstract – The status quo in visual recognition is to learn from “disembodied” bags of labeled snapshots from the Web. Yet cognitive science tells us that perception develops in the context of acting and moving in the world—and without intensive supervision. This discrepancy prompts the question: How can unlabeled video augment visual learning? I’ll describe our recent work exploring how a system can learn effective representations by watching unlabeled video. First, we show how 3D egomotion signals accompanying the video provide a valuable cue during learning, allowing the system to internalize the link between “how I move” and “what I see”. Building on this idea, we develop a generalization of slow feature analysis that captures higher order temporal coherence. We demonstrate the methods’ impact for various recognition tasks. This includes scenarios for active recognition, where the agent intelligently moves to disambiguate its surroundings, as well as traditional image classification tasks. For the latter, we show that features learned from ego-video on an autonomous car substantially improve large-scale scene recognition, while those learned from unlabeled YouTube videos can even surpass a standard heavily supervised CNN pretraining approach.
Bio – Kristen Grauman is an Associate Professor in the Department of Computer Science at the University of Texas at Austin. Her research in computer vision and machine learning focuses on visual search and object recognition. Before joining UT-Austin in 2007, she received her Ph.D. in the EECS department at MIT, in the Computer Science and Artificial Intelligence Laboratory. She is an Alfred P. Sloan Research Fellow and Microsoft Research New Faculty Fellow, a recipient of NSF CAREER and ONR Young Investigator awards, the Regents' Outstanding Teaching Award from the University of Texas System in 2012, the PAMI Young Researcher Award in 2013, the 2013 Computers and Thought Award from the International Joint Conference on Artificial Intelligence, and a Presidential Early Career Award for Scientists and Engineers (PECASE) in 2013. She and her collaborators were recognized with the CVPR Best Student Paper Award in 2008 for their work on hashing algorithms for large-scale image retrieval, and the Marr Best Paper Prize at ICCV in 2011 for their work on modeling relative visual attributes. She serves on the Editorial Board for the International Journal of Computer Vision(IJCV), as an Associate Editor in Chief for the Transactions on Pattern Analysis and Machine Intelligence (TPAMI), and served as a Program Chair of CVPR 2015 in Boston.Eitan Grinspun, Columbia University, USA - The Geometry of Physics Has a Profound Impact on ComputationBio – I am an Associate Professor of Computer Science and Applied Mathematics at Columbia University, where I conduct research at the intersection of geometry, physics, and computation. I am particularly interested in discrete differential geometry and its applications to computational mechanics. I was Professeur d'Université Invité in Paris at l'Université Pierre et Marie Curie in 2009. I completed my postdoc at the Courant Institute of Mathematical Sciences in 2004, my graduate studies at Caltech in 2003, and my undergraduate in Engineering Science at the University of Toronto in 1997. My work was recognized by an Alfred P. Sloan Research Fellowship, NSF CAREER Award, and NVIDIA Fellowship. I was the Program co-chair for the ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA’09) in 2009, the Symposium on Geometry Processing (SGP’12) in 2012, and Pacific Graphics in 2016. I serve as Associate Editor for ACM Transactions on Graphics.
The technologies developed with my collaborators are used in Adobe Photoshop/Illustrator, and at film studies such as Disney, Pixar, Blue Sky and Weta Digital. My work was profiled in the New York Times, Scientific American, Popular Science (“Brilliant Ten Scientists of 2011”), and Fast Company Magazine (“Most Creative People of 2013”).Leonid Karlinsky, IBM Research, Israel - Fine-Grained Recognition of Thousands of Object Categories with Single-Example TrainingAbstract – We approach the problem of fast detection and recognition of a large number (thousands) of object categories while training on a very limited amount of examples (usually one) per category. Examples of this task include: (i) detection of retail products, where we have only one studio image of each product available for training; and (ii) detection of 3D objects and their respective poses within a single monocular image, where only a sparse subset of (partial) object views is available for training, with a single example for each view. Building a detector based on such a small amount of examples presents a significant challenge for the current top-performing (deep) learning based techniques, which require large amounts of data to train. Our approach for this task is based on a non-parametric probabilistic model for initial detection, CNN-based refinement and temporal integration where applicable. We successfully demonstrate its usefulness in a variety of experiments on both existing and our own benchmarks showing performance that compares favorably to the current state-of-the-art
Bio –Dr. Karlinsky leads the AR research in the Computer Vision and Interaction (CVI) group @ IBM Research. He is a Computer Vision and Machine Learning expert with years of hands on experience. He has published research papers in leading CV and ML venues such as ECCV, CVPR and NIPS and is actively reviewing for these conferences for the past 7 years. Dr. Karlinsky holds a PhD degree in CV from the Weizmann Institute of Science, supervised by Prof. Shimon Ullman.Michael Kazhdan, Johns Hopkins University, USA - Signal Processing - from Images to SurfacesAbstract – In this talk, we will review a number of well-established techniques for processing images over regular grids and describe how they can be extended to the processing of images and signals over meshes. We will start by considering simple (linear) techniques such as image stitching and blurring/sharpening and then proceed to more complex (non-linear) techniques like shock filtering and optical flow. Though the applications differ, we will show how their extension to meshes rely on the same underlying geometry-processing tools.
Bio – Michael Kazhdan is an associate professor in the Computer Science Department at Johns Hopkins University. Some of his recent research focuses on the challenge of translating established techniques for image processing into techniques for processing signals over meshes. These have included simpler linear operations such as compositing and filtering as well as non-linear operations such shock filtering and optical flow.Ira Kemelmacher-Shlizerman, University of Washington, USA - People Modeling from Historical FootageAbstract – I will report on my group's venture intro modeling people from massive amounts of Internet photos and videos.
Bio – Ira Kemelmacher-Shlizerman is an Assistant Professor in the Department of Computer Science and Engineering at the University of Washington. She received her Ph.D in computer science and applied mathematics at the Weizmann Institute of Science. Ira works in computer vision and computer graphics, and specifically on 3D reconstruction and modeling of people from big uncalibrated photo and video collections (Internet, personal photos, etc.) Key applications are in virtual and augmented reality and visualizations of massive photo collections.
Ira received the Google faculty award, her work “Moving Portraits” was selected to the cover of the Communications of the ACM, Research Highlights, and tech transferred to Google. Her work “Illumination aware age progression” and its application to missing children search was featured by interviews on national TV, e.g., CBS, NBC, and many others. Ira's 3D face reconstruction from Internet photos received the Madrona prize, and the "Innovation of the 2016 Year Award" by Geekwire.Mark Kliger, Intel, Israel - Taking a Deeper Look at 3D ScenesAbstract – Since the early days of computer vision the problem of 3D indoor scene understanding has been considered one of its "holy grails". To tackle this problem we need to have an ability to jointly solve multiple challenging computer vision problems: 3D object detection, semantic segmentation, pose estimation, object retrieval and registration, etc. In the past several years we have witnessed significant progress towards solutions of some of these problems, partially due to the successes of Deep Learning and the appearance of consumer-grade 3D cameras. In this talk, we will discuss current work within Intel’s Perceptual Computing group on 3D indoor scene understanding, based on Intel's RealSense 3D camera and using modern algorithmic methods.
Bio – Mark Kliger is a research scientist at Intel Perceptual Computing group, working on computer vision and machine learning problems. Prior to Intel, Mark and his colleagues developed gesture recognition software at Omek Interactive, which was acquired by Intel in 2013. Before Omek Mark was the CTO of Medasense Biometrics, where he led the development of pain assessment technology. Mark obtained his Ph.D. in statistical signal processing from Ben-Gurion University in 2006 and from 2006 to 2008 was a research fellow at the University of Michigan, Ann-Arbor, where he worked on machine learning problemsYaron Lipman, Weizmann Institute of Science, Israel - Drawing Graphs in OrbifoldsAbstract –Tutte’s graph drawing algorithm is one of the most popular techniques for computing parameterizations of surface meshes into the plane as it guarantees validity, simple to implement and compute, and minimizes a natural distortion energy.
So far, Tutte's algorithm was applicable to two types of target domains: convex polygonal planar domains and flat tori. In this talk we show how Tutte's algorithm can be generalized to handle a richer set of target domains called Orbifolds. Orbifolds are simple surfaces exhibiting different topologies and cone singularities, defined using symmetry groups.
We demonstrate two applications made possible by the Orbifold-Tutte algorithm: Linear approximation of conformal maps, and homeomorphic surface to surface mappings.
Guy Medan, BrightSource, Israel - Solar VisionAbstract – Concentrated Solar Power Plants (CSP) transform the sun's energy into electricity using a large array of mirrors to produce steam, which drives a conventional turbine. The Power Tower design uses a receiver atop a central tower and flat, computer-controlled mirrors to concentrate the sun's rays onto it. Brightsource Energy designs, develops and deploys Solar Power Tower plants, with a dedicated control system which relies on many sensors, including cameras and computer vision algorithms, for achieving optimal power generation under dynamic conditions. Such algorithms are used for various tasks in the solar field including mirror production QC, calibration, feedback and prediction. In the talk several challenges in these areas and their solutions will be presented.
Bio – Guy Medan is an algorithms developer at Brightsource Modelling group, working on machine vision and optimization problems, and a PhD candidate at the Medical Image Processing laboratory, HUJI School of Computer Science and Engineering. Guy received his B.Sc in EE & Physics from the Technion.Niloy Mitra, University College London, UK - Computational Design of Functional ObjectsAbstract –Both professionals and hobbyists like to design functional objects for physical use. However, there exists limited computational support to facilitate such a design process. Existing tools either require specialized skills and extensive training, or necessitate users to perform extensive trial and error based exploration with limited guidance. In this talk, I will discuss some of our recent attempts to computationally create non-trivial design variations and show them to the users to help them explore and pick novel design alternatives. I will present specific example cases with folded paper models and also design patterns. More details can be found at: http://geometry.cs.ucl.ac.uk/.
Bio – Niloy J. Mitra leads the Smart Geometry Processing group in the Department of Computer Science at University College London (UCL). He received his PhD degree from Stanford University under the guidance of Leonidas Guibas. His research interests include shape understanding, computational design, geometry processing, and more broadly computer graphics. Niloy received the ACM Siggraph Significant New Researcher Award in 2013 for his work on 'discovery and use of structure in 3D objects' and the BCS Roger Needham Award in 2015. His work has twice been featured as research highlights in Communication of ACM. Besides research, Niloy is an active DIYer and loves reading, climbing, and cooking. For more, please visit: http://geometry.cs.ucl.ac.uk/Srinivasa Narasimhan, Carnegie Mellon University, USASrinivasa Narasimhan is an Associate Professor in the Robotics Institute at Carnegie Mellon University. He obtained his PhD from Columbia University in Dec 2003. His group focuses on novel techniques for imaging, illumination and light transport to enable applications in vision, graphics, robotics, agriculture and medical imaging. His works have received several awards: Best Demo Award (IEEE ICCP 2015), A9 Best Demo Award (IEEE CVPR 2015), Marr Prize Honorable Mention Award (2013), FORD URP Award (2013), Best Paper Runner up Prize (ACM I3D 2013), Best Paper Honorable Mention Award (IEEE ICCP 2012), Best Paper Award (IEEE PROCAMS 2009), the Okawa Research Grant (2009), the NSF CAREER Award (2007), Adobe Best Paper Award (IEEE Workshop on Physics based methods in computer vision, ICCV 2007) and IEEE Best Paper Honorable Mention Award (IEEE CVPR 2000). His research has been covered in popular press including NY Times, BBC, PC magazine and IEEE Spectrum and is highlighted by NSF and NAE. He is the co-inventor of programmable headlights, Aqualux 3D display, Assorted-pixels, Motion-aware cameras and Episcan3D. He co-chaired the International Symposium on Volumetric Scattering in Vision and Graphics in 2007, the IEEE Workshop on Projector-Camera Systems (PROCAMS) in 2010, and he IEEE International Conference on Computational Photography (ICCP) in 2011, co-edited a special journal issue on Computational Photography, and serves on the editorial board of the International Journal of Computer Vision and as Area Chair of top computer vision conferences (CVPR, ICCV, ECCV, BMVC, ACCV).Frédéric Parienté, NVIDIA - Advances in GPU Architecture for Deep Learning and Scientific ComputingAbstract – The talk will cover the recent NVIDIA product announcements made at the GTC'16 conference, and how the Pascal GPU and NVLink interconnect technologies greatly improve multi-GPU performance and efficiency in deep learning and scientific computing applications.
Bio –Frédéric Parienté is a Business Development Manager for Accelerated Computing at NVIDIA since 2015, in charge of Higher Education and Research across Southern Europe. Previously, he spent his engineering career at Sun Microsystems–Oracle as a performance software engineer and was the regional Director of ISV Engineering when he left Oracle in 2015. Frédéric graduated in General Engineering from ENSTA ParisTech, Mechanical Engineering from University of Illinois and Finance from Université Paris Dauphine.Marc Pollefeys, ETH Zurich, Switzerland - Semantic 3D Reconstruction
Abstract –While purely geometric models of the world can be sufficient for some applications, there are also many application that need additional semantic information. In this talk I will focus on 3D reconstruction approaches which combine geometric and appearance cues to obtain semantic 3D reconstructions. Specifically, the approaches I will discuss are formulated as multi-label volumetric segmentation, i.e. each voxel gets assigned a label corresponding to one of the semantic classes considered, including free-space. We propose a formulation representing raw geometric and appearance data as unary or high-order (pixel-ray) energy terms on voxels, with class-pair-specific learned anisotropic smoothness terms to regularize the results. We will see how by solving both reconstruction and segmentation/recognition jointly the quality of the results improves significantly and we can make progress towards 3D scene understanding.
Bio – Marc Pollefeys is a full professor and head of the Institute for Visual Computing of the Dept. of Computer Science of ETH Zurich which he joined in 2007. He leads the Computer Vision and Geometry lab. Previously he was on the faculty at the University of North Carolina at Chapel Hill. HE obtained his MS and PhD degrees from the KU Leuven in Belgium. His main area of research is computer vision. One of his main research goals is to develop flexible approaches to capture visual representations of real world objects, scenes and events. Dr. Pollefeys has received several prizes for his research, including a Marr prize, an NSF CAREER award, a Packard Fellowship and a ERC Starting Grant. He is the author or co-author of more than 280 peer-reviewed papers. He will be general chair of ICCV 2019, was a general chair for ECCV 2014 in Zurich and one of the program chairs for the IEEE Conf. on Computer Vision and Pattern Recognition 2009. Prof. Pollefeys was on the Editorial Board of the IEEE Transactions on Pattern Analysis and Machine Intelligence, the International Journal of Computer Vision, Foundations and Trends in Computer Graphics and Computer Vision and several other journals. He is an IEEE Fellow.Helmut Pottmann, TU Wien, Austria - Modeling Architectural FreeformAbstract – Architectural projects of high geometric complexity greatly benefit from the incorporation of essential aspects of function, fabrication and statics into the shape modeling process. This integrated view is one of the major goals of Architectural Geometry. One has to avoid detailed physical simulation as this would hardly be compatible with interactive shape manipulation. Instead, we aim at developing shape modeling tools which employ simplified mathematical models in order to respect manufacturing and structural constraints. In the present talk, we will focus on architectural form-finding problems, in particular on self-supporting structures. We access them through meshes in force equilibrium and discuss their connections to discrete differential geometry. Interactive modeling is enabled through an efficient projection onto constraint varieties whose algebraic structure is simplified by introducing auxiliary variables and ensuring that constraints are at most quadratic. This is joint work with Chengcheng Tang, Etienne Vouga, Mathias Höbinger, Alexandra Gomes, Xiang Sun and Johannes Wallner.
Bio –Helmut Pottmann is a professor of Applied Geometry and director of the Center for Geometry and Computational Design at Vienna University of Technology. From 2009-2013, he has been director of the Geometric Modeling and Scientific Visualization Center at King Abdullah University of Science and Technology. His research interests are in Applied Geometry and Visual Computing, in particular in Geometric Modeling, Geometry Processing and most recently in Geometric Computing for Architecture and Manufacturing. He has co-authored two books and more than 200 refereed articles.Tali Treibitz, University of Haifa, Israel - 3D in the OceanAbstract – The ocean covers 70% of the earth surface, and influences almost every aspect in our life. It is a complex foreign environment that is hard to explore and therefore much about it is still unknown. As human access to most of the ocean is very limited, novel imaging systems and computer vision methods are needed to reveal new information about the ocean that is currently unknown. However, the environment poses numerous challenges on reconstruction, such as handling optics through a medium, movement, and limited resources, while operating in a large-scale environment. In the talk I will give an overview of current efforts and challenges in this field.
Bio – Tali Treibitz is heading the Marine Imaging Lab in the School of Marine Sciences in the University of Haifa. She received the BA degree in computer science and the PhD degree in electrical engineering from the Technion-Israel Institute of Technology in 2001 and 2010, respectively. Between 2010-2013 she has been a post-doctoral researcher in the department of computer science and engineering, in the University of California, San Diego and in the Marine Physical Lab in Scripps Institution of Oceanography.Yair Weiss, The Hebrew University of Jerusalem, Israel - Statistics of RGBD ImagesAbstract – Cameras that can measure the depth of each pixel in addition to its color have become easily available and are used in many consumer products worldwide. Often the depth channel is captured at lower quality compared to the RGB channels and different algorithms have been proposed to improve the quality of the D channel given the RGB channels. Typically these approaches work by assuming that edges in RGB are correlated with edges in D. In this paper we approach this problem from the standpoint of natural image statistics. We obtain examples of high quality RGBD images from a computer graphics generated movie (MPI-Sintel) and we use these examples to compare different probabilistic generative models of RGBD image patches. We then use the generative models together with a degradation model and obtain a Bayes Least Squares (BLS) estimator of the D channel given the RGB channels. Our results show that learned generative models improve the state-of-the-art in improving the quality of depth channels given the color channels in natural images even when training is performed on artificially generated images.