Title of the Talk: Building a Visual Computing Career in Industrial Research Labs

Date&Time: 19 December 2020, 5:30 PM (IST)
Session Chair: P J Narayanan

Abstract: In this talk I'll share my experience in building a visual computing career in industrial research labs. The conventional wisdom in academia is to choose a specific area, also known as "your niche", and try to establish yourself in it. This wisdom is further reenforced by the social proof of the papers selection processes at research conferences. An alternative approach, which often works better, is to break the traditional boundaries of different areas and focus on solving important problems as they naturally arise in the real world. Essentially you end up devoting your career to a few problems that you're deeply passionate about and to those you believe will have big intellectual and social impact, regardless where these problems are positioned with respect to traditional academic boundaries. This different approach might look more risky but it could be more rewarding and has become attractive even to academic researchers lately as various visual computing fields become increasingly mature and interdisciplinary. I'll present two research projects, one in deep-fake detection and the other in texture modeling, as case studies to illustrate some of the finer points on how to maximize impact through long-term thinking.

Baining Guo is a Distinguished Scientist with Microsoft Corporation and Assistant Managing Director of Microsoft Research Asia, where he also serves as the head of the computer graphics lab. Prior to joining Microsoft in 1999, Baining was a senior staff researcher with Intel Research in Santa Clara, California. Baining received his PhD and MS degrees from Cornell University, and his BS from Beijing University. He is a fellow of ACM, IEEE, and Canadian Academy of Engineering.
Baining works in computer graphics and computer vision. His interests span most aspects of computer graphics, with an emphasis on statistical modeling of texture and appearance, GPU-based rendering, and geometric modeling. He has also worked on image understanding and video analysis. He was on the editorial boards of IEEE Transactions on Visualization and Computer Graphics, Elsevier Journal of Computer and Graphics, and IEEE Computer Graphics and Applications. He has also served on program committees of most major graphics and visualization conferences, including ACM SIGGRAPH, ACM SIGGRAPH Asia, and IEEE Visualization. In 2014, he was the technical papers chair of ACM SIGGRAPH Asia. Dr. Guo has over 50 US patents.

Title of the talk: Building Interpretable Machine Learning & Learning Robust Model Fitting

Date&Time: 20 December 2020, 2:00 PM (IST)
Session Chair: Subhashis Banerjee

Abstract: In this talk I will introduce two lines of research which we have been conducting in our lab in the last years. The first one considers a specific type of neural network which is per-construction invertible, also known as normalizing flows. We look at various advantages of this network architecture, in contrast to standard feed-forward networks. In particular, multiple diverse outputs, uncertainty estimation, and competitive generative classification. The second line of research considers the task of fitting low-dimensional parametric models to data. Examples are 6D camera localization for SLAM, horizon line detection, or Essential Matrix estimation. We show that it is worth to combine traditional, robust methods, such as RANSAC, with Neural Networks. Doing so, leads to networks that generalize better, and to networks that are allowed to make gross mistakes.

Carsten Rother received the diploma degree with distinction in 1999 from the University of Karlsruhe/Germany, conducting his diploma thesis with Prof. Dr. H.-H. Nagel. He received his PhD degree in 2003 from the Royal Institute of Technology Stockholm/Sweden, under the guidance of Jan-Olof Eklundh and Stefan Carlsson. From 2003 until 2013 he was researcher with Microsoft Research Cambridge/UK, and a member of the Computer Vision Group lead by Andrew Blake. From 2014 until 2017 he was full (W3) Professor at TU Dresden. Since September 2017 he is full Professor at Uni Heidelberg, heading the Visual Learning Lab Heidelberg. He is also coordinating director of the Heidelberg Collaboratory for Image Processing (HCI) 3rd phase. His research interests are in the field of computer vision and machine learning - ranging from deep learning and graphical models to smart data generation. He has been working on a broad range of applications - such as image editing (e.g. interactive image segmentation, alpha matting, and deconvolution), image matching (e.g. large displacement Scene Flow), scene understanding (e.g. 6D object pose estimation), Bio-Imaging (e.g. cell tracking). He has published over 150 articles (current H-index 60) at international conferences and journals. He won awards at BMVC'16, ACCV'14, CVPR'13, BMVC'12, ACCV'10, CHI'07, CVPR'05, and Indian Conference on Computer Vision'10. He was awarded the DAGM Olympus prize in 2009. He has co-developed two Microsoft products, GrabCut for Office 2010 and AutoCollage. He also co-authored a book on Markov Random Fields in Computer Vision and Image Processing. He serves as area chair for major conferences and he has been associated editor for T-PAMI.

Title of the Talk: Towards making AI a commodity - Closing the Gap Between Research and Practice

Date&Time: 21 December 2020, 2:00 PM (IST)
Session Chair: Santanu Chaudhury

Abstract: Research in AI is booming! The numbers of algorithms being published and companies that invest tremendous resources into AI research are constantly on the rise. Yet, it is still challenging to translate research in Computer Vision into practical working products that are used by millions of people. In this talk I will approach this question through the lens of Photo Drive applications, which require solutions for some of the core problems in computer vision, such as image classification and search. While there are many published algorithms for these tasks, many of them are not fit for real-world scenarios. Some of the reasons for this gap is that in real-life, solutions need to align to the product user experience requirements as well as comply with multiple objectives, such as budgeted resources and limited data. In this talk, I will introduce our recent work on efficient image understanding, designed to be used in practice in Photo Drive apps, discussing use-cases such as image grouping based on visual content and image search through free-text query. I will share our proposed approach to solving them, such as employing neural architecture search, advanced training techniques and more.

Lihi Zelnik-Manor is an Associate Professor in the Faculty of Electrical Engineering at the Technion and the Head of Alibaba DAMO Academy Machine Intelligence Israel Lab. Prof. Zelnik-Manor holds a PhD and MSc (with honors) in Computer Science from the Weizmann Institute of Science and a BSc (summa cum laude) in Mechanical Engineering from the Technion. Her main area of expertise is Computer Vision. Prof Zelnik-Manor has done extensive community contribution, serving as Program Chair of CVPR'16, Associate Editor at TPAMI, served multiple times as Area Chair at CVPR, ECCV and was on the award committee of ACCV'18 and CVPR'19. Looking forward she will serve as General Chair of CVPR'21 and ECCV'22 and as Program Chair of ICCV'25.