Plenary_Speakers

Wolfgang Heidrich

Prof. Wolfgang Heidrich is a Professor of Computer Science and the Director of the Visual Computing Center at King Abdullah University of Science and Technology (KAUST). Prof. Heidrich joined KAUST in 2014, after 13 years as a faculty member at the University of British Columbia. He received his Ph.D. from the University of Erlangen in 1999, and then worked as a Research Associate in the Computer Graphics Group of the Max-Planck-Institute for Computer Science in Saarbrucken, Germany, before joining UBC in 2000. Prof. Heidrich’s research interests lie at the intersection of imaging, optics, computer vision, computer graphics, and inverse problems. His more recent interest is in computational imaging, focusing on hardware-software co-design of the next generation of imaging systems, with applications such as High-Dynamic Range imaging, compact computational cameras, hyperspectral cameras, to name just a few. Prof. Heidrich’s work on High Dynamic Range Displays served as the basis for the technology behind Brightside Technologies, which was acquired by Dolby in 2007. Prof. Heidrich is a Fellow of the IEEE and Eurographics and the recipient of a Humboldt Research Award.

Deep Optics — Joint Design of Imaging Hardware and Reconstruction Methods

Classical imaging systems are characterized by the independent design of optics, sensors, and image processing algorithms. In contrast, computational imaging systems are based on a joint design of two or more of these components, which allows for greater flexibility of the type of captured information beyond classical 2D photos, as well as for new form factors and domain-specific imaging systems. In this talk, I will describe how numerical optimization and learning-based methods can be used to achieve truly end-to-end optimized imaging systems that outperform classical solutions.

Abdesselam Bouzerdoum

Abdesselam Bouzerdoum (M’89–SM’03 IEEE) graduated with MSEE and Ph.D. degrees from the University of Washington, Seattle, USA. In 1991, he joined Adelaide University, South Australian, and from 1998 to 2004 he was an Associate Professor with Edith Cowan University, Perth, Western Australia. In September 2004, he was appointed Professor of Computer Engineering and Head of School of Electrical, Computer & Telecommunications Engineering at the University of Wollongong. From 2007 to 2013, he served as Associate Dean (Research), Faculty of Informatics. Since March 2017, he has also been a Professor with Hamad Bin Khalifa University (HBKU), Qatar. Dr. Bouzerdoum held several Visiting Professor Appointments at Institut Galilée, Université Paris-13, LAAS/CNRS, Toulouse, France, Villanova University, USA, and the Hong Kong University of Science and Technology. From 2009 to 2011, he was a member of the ARC College of Experts and Deputy Chair of the EMI panel (2010–2011).Dr. Bouzerdoum is the recipient of the Eureka Prize for Outstanding Science in Support of Defence or National Security (2011), the Chester Sall Award of IEEE Trans. Consumer Electronics (2005), and a Distinguished Researcher Award (Chercheur de Haut Niveau) from the French Ministry (2001). He served as Associate Editor for 5 international journals, including IEEE Trans. Image Processing, IEEE Trans. Systems, Man, and Cybernetics (1999–2006). He has published over 350 technical articles and graduated over 30 Ph.D. and Research Masters students. His research interest includes radar imaging and signal processing, image processing, vision, machine learning, and pattern recognition.

Semantic Segmentation for Assistive Navigation using Deep Bayesian Gabor Networks

Semantic scene segmentation is a challenging problem that has great importance in assistive and autonomous navigation systems. Such vision systems must cope well with image distortions, lighting variations, changing surfaces, and varying illumination conditions. For a vision-impaired person, the task of navigating in an unstructured environment presents major challenges and constant danger. It is reported that on average one in 12 pedestrians living with blindness is hit by a cyclist, a motorbike, or a car. Safe navigation involves multiple cognitive tasks at both macro and micro levels. At the macro level, a blind person needs to know the general route to take, his/her location along the route at any time, and the relevant landmarks and intersections. At the micro-level, the blind person needs to walk within the pedestrian lane on a safe surface, maintain his or her balance, detect obstacles in the scene, and avoid hazardous situations. To support the vision impaired navigating safely in unconstrained outdoor environments, an assistive vision system should perform several vital tasks such as finding pedestrian lanes, detecting and recognizing traffic obstacles, and sensing dangerous traffic situations.

In this talk, we will present vision-based assistive navigation systems that can segment objects in the scene, measure their distances, identify pedestrians, and detect a walking path. Using range and intensity images enable fast and accurate object segmentation and provide useful navigation cues such as distances to nearby objects and types of objects. Furthermore, the talk will present a new hybrid deep learning approach for semantic segmentation. The new architecture combines Bayesian learning with deep Gabor convolutional neural networks (GCNNs) to perform semantic segmentation of unstructured scenes. In this approach, the Gabor filter parameters are modeled as normal distributions with mean and variance that are learned using variational Bayesian inference. The resulting network has a smaller number of trainable parameters, which helps mitigate the overfitting problem while maintaining the modeling power. In addition to the output segmentation map, the system provides two maps of aleatoric and epistemic uncertainty—a measure that is negatively correlated with the confidence level with which we can trust the segmentation results. This measure is important for assistive navigation applications since its prediction affects the safety of its users. Compared to the state-of-the-art semantic segmentation methods, the hybrid Bayesian GCNN yields a competitive segmentation performance with a very compact architecture (a size reduction of between 25.4 and 231.2 times), a fast prediction time (1.6 to 67.4 times faster), and a well-calibrated uncertainty measure.

Stéphane Cotin

Stéphane Cotin is the Research Director at Inria and the leader of the MIMESIS team. His main research interests are in the physics-based simulation, real-time simulation of soft-tissue deformations, finite element modeling, augmented reality and medical applications of this research. In an effort to disseminate their work and to accelerate research in these fields, he initiated the SOFA project (www.sofa-framework.org) which is now a reference Open Source solution for developing advanced simulations. He is the author or co-author of more than 150 scientific articles and involved with several conferences, either as organizer or member of the board: International Symposium on Biomedical Simulation (ISBMS), International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), International Conference on Information Processing in Computer-Assisted Interventions (IPCAI).
Since 2012 he is also a co-founder and president of the scientific board of InSimo (www.insimo.fr). Before joining Inria, he was Research Lead for the Medical Simulation Group at CIMIT and an instructor at Harvard Medical School.

Augmented Surgery: how numerical simulation, computer vision and machine learning can help reduce risks in the operating room

Developments of imaging devices, numerical methods, and medical robotics are profoundly changing how modern medicine is practiced. This talk will highlight the increasing role of real-time numerical simulation in the fields of surgery and interventional radiology, with an impact in two major application areas: surgical training and computer-assisted interventions.

Numerical simulations have been used for several decades to perform complex biomedical phenomena analysis, with various levels of success. A specificity of our application contexts is the need for real-time simulations adapted to each patient. To this end, we have developed specific numerical methods, which allow for real-time computation of finite element simulations. While information about the organ shape, for instance, can be obtained pre-operatively, other patient-specific parameters can only be determined intra-operatively. This is achieved by exploiting our application domain's context, where images of different nature are acquired during the intervention. Machine learning methods are then used to extract information from these images and, in some cases, to replace the numerical simulation itself. An illustration of these different topics will be demonstrated by modeling liver biomechanics and its parametrization to achieve patient-specific augmented reality during surgery.

Pascal Mamassian

Pascal Mamassian was born in Lyon (France) and studied in Telecommunication Engineering (Sup' Télécom, Paris, France). He then obtained a Master's degree in Cognitive Sciences (Univ. Paris 6 & EHESS, Paris, France) and a Ph.D. degree in Experimental and Biological Psychology (Univ. of Minnesota, Minneapolis, USA). He has worked at the Max-Planck Institute for Biological Cybernetics (Tübingen, Germany) and New York University (New York, USA). He was at the University of Glasgow (Glasgow, UK) as a lecturer and then senior lecturer before taking a researcher position at the CNRS (France). Since 2014, he is the founding director of the Laboratoire des Systèmes Perceptifs at the Ecole Normale Supérieur In Paris.

Visual Confidence

Visual confidence refers to our ability to predict the correctness of our perceptual decisions. Knowing the limits of this ability, both in terms of biases (e.g. overconfidence) and sensitivity (e.g. blindsight), is clearly important to approach a full picture of perceptual decision making. In recent years, we have explored visual confidence using a paradigm called confidence forced-choice. In this paradigm, observers have to choose which of two perceptual decisions is more likely to be correct. I will review some behavioral results obtained with the confidence forced-choice paradigm, together with a theoretical model based on signal detection theory.