Program

Correctly matching feature points in a pair of images is an important pre-processing step for many computer vision applications. In this talk we propose an efficient method for estimating the number of correct matches without explicitly computing them. In addition, our method estimates the region of overlap between the images. To this end, we propose to analyze the set of matches using the spatial order of the features, as projected to the x-axis of the image. The set of features in each image is thus represented by a sequence. This reduces the analysis of the matching problem to the analysis of the permutation between the sequences.

Using the Kendall distance metric between permutations and natural assumptions on the distribution of the correct and incorrect matches, we show how to estimate the above-mentioned values. We demonstrate the usefulness of our method in two applications: (i) a new halting condition for RANSAC based epipolar geometry estimation methods that considerably reduce the running time, and (ii) discarding spatially unrelated image pairs in the Structure-from-Motion pipeline. 

Authors: Lior Talker, Yael Moses, Ilan Shimshoni

Ilan Shimshoni received his B.Sc. in mathematics from the Hebrew University in Jerusalem, his M.Sc. in computer science from the Weizmann Institute of Science, and his Ph.D. in computer science from the University of Illinois at Urbana Champaign (UIUC).

Ilan was a post-doctorate fellow at the faculty of computer science at the Technion, from 1995-1998, and was a member of the faculty of industrial engineering and management from 1998-2005. He joined the department of  Information Systems (IS) at Haifa University in October 2005.

His research areas are computer vision, computer graphics and their application to for example archaeology. He is an associate editor of IEEE PAMI.

Abstract

While augmented reality (AR) has been steadily gaining popularity in recent years, the majority of AR solutions are either mixed reality (MR), where graphics are not attached to specific elements in the scene, or marker-based AR, where graphics are anchored to a planar object visible in the scene. These solutions are often suitable for limited scenarios, but typical use-cases in the industrial domain require precisely anchored, marker-free AR. Our work in the Cognitive Vision and Augmented Reality group at IBM Research in Haifa focuses on research in AR for technical training, field maintenance, IoT, and other use-cases. In this talk I will present the group's AR platform and review some of the real-world installations in the industrial domain.

Yochay Tzur leads the 3D vision and augmented reality research at IBM Research - Haifa. He joined IBM in 2013, with over 10 years of hands-on experience in companies such as Rafael and CSR-Zoran. At IBM, Yochay has been leading the development of AR projects, from prototype stage to realization at the customer's site. In his current role, he is responsible for the development of the complete AR offering of the lab. Yochay received his B.Sc. and M.Sc degrees in computer science from the Technion and is the author of several patents.

Abstract

VR headsets use lenses to expand their field of view and to enable the viewer’s eyes to focus on the screen. However, the lenses introduce distortion in the image. This requires the graphic engine to render an image that is distorted in the opposite way, over-shading the outskirts of the image, rendering and shading a lot of pixels that will not make it to the displayed image. This is a significant inefficiency.
This talk will present two techniques for overcoming this inefficiency: Multi-Resolution Shading and Lens Matched Shading. We will also talk about NVIDIA's VRWorks and how it helps AR/VR developers.

Yaki Tebeka is a Distinguished Engineer at NVIDIA, located at the newly established NVIDIA Tel Aviv R&D Center, where he creates best-in-class software development tools, to help developers write robust and optimized code for NVIDIA's platforms and solutions. Yaki brings two decades of experience in software development, focusing on 3D graphics, heterogeneous computing and developer tools. Before joining NVIDIA, Yaki worked at AMD and was a co-founder of Graphics Remedy, acquired by AMD in 2010.
Yaki received his B.Sc. degree in computer science and mathematics from the Bar Ilan University

Abstract

Deep learning plays a key role in the race to autonomous vehicles. Although remarkable progress has been made, the vast majority of both existing theories and technologies have yet to transition to real-world scenarios, which introduce a huge variety of road, weather and lighting conditions as well as deviations of driver behavior.  
Nexar builds the world's largest open vehicle-to-vehicle (V2V) network by turning smartphones into connected AI dash-cams. Joining deep learning with millions of crowdsourced driving miles collected by our users, Nexar’s technology provides a new, safer driving experience with the potential of saving the lives of 1.3 million people who die on the road every year.
In this talk, we will share our journey to make the roads safer. We will examine some of the challenges we face, from a real-time collision avoidance system to learning autonomous driving policies for a safer driving experience. We will also share some of our joint research with the Berkeley Deep-Drive (BDD) Industry Consortium and present the Nexar challenge, in which we open some of our deep learning challenges to the outside world and invite aspiring researchers to test their chops, win prizes, and to join our mission to free the world of car accidents.

Ilan Kadar is the Director of Deep Learning at Nexar. Ilan is responsible for leading the deep learning team and effort to leverage Nexar's large-scale datasets of real-world driving environments to automative applications. Prior to Nexar, Ilan was leading the deep learning group at Cortica and was responsible for building the company's machine vision technology. Ilan received his BSc, MSc and PhD degrees in computer science from the Ben-Gurion University of the Negev, Israel, in 2006, 2008, and 2012 respectively (Summa Cum Laude). His research thesis focused on machine learning algorithms for scene recognition and image retrieval, while employing insights from behavioral and psychophysical experiments. His work was published in leading conferences and journals in the areas of machine vision and was awarded the best research project at IMVC in 2013, the Intel award for excellent Israeli PhD students in 2012, and the Friedman award for outstanding PhD students in 2012.

Abstract

A revolution has started…
Robots are traditionally confined to factories behind steel fences for the safety of the workers around them. That is changing thanks to advanced perception capabilities, robots are now getting ready to be a part of our day to day lives.
For a researcher, joining this revolution is not a simple task. There is a steep learning curve to get into robotics. It is complicated, and it involves working with many devices and different technologies. What if creating perceptive robots would have been easy and fun? Just plug and play. What would you do?
Join us to hear more about Intel® Euclid™ - an innovative all-in-one perception device that allows beginners and professionals to use advanced vision capabilities in their robotic projects.

Amit Moran is leading the Robotics Innovation team as part of the Intel's RealSense Innovation Lab in Israel. In the past 5.5 years in the Innovation Lab, Amit has lead several novel experiences and usages, leveraging from Intel’s RealSense™ program; during which he has been inventing, designing and prototyping innovative concepts for interacting with our computing devices. Currently, Amit is leading a R&D team focused on 3D vision in robotics. Amit holds a MSc degree in Software Engineering from Institute National de Science Appliqué (INSA) de Lyon, France. He is a co-author of several patents in the perceptual computing (NUI) domain.

Abstract

Reconstruction of a full dynamic scene using a monocular moving camera requires reconstruction of both static and dynamic objects. We present a generic approach for simultaneous estimation of shape, location and velocity for static and moving objects using graph optimization. The main challenge is the unknown scale of each moving object; this is handled with a good initial depth estimation. For validation we created a dynamic simulation of an urban environment. On this simulation our method yields 13% velocity error when the camera is mounted on a car, and 2.8% velocity error when it is mounted on a drone.

I received my B.Sc in computer engineering from the Hebrew University, specializing in electro-optics and microelectronics. In 2012 I received my M.Sc in applied physics, specializing in nanotechnology. I currently work at Rafael's Jerusalem site as a computer vision algorithm developer.

Abstract

This talk will introduce the listeners to the world of capsule endoscopy and its computer vision and machine learning challenges. One application of this pioneering technology is as a screening tool for detecting polyps, a precursor to colorectal cancer. 
We will go over some of the technological innovations which are part of our research.  The highlight of the talk will introduce a novel and general-purpose classification algorithm aimed at directly optimizing for the image classification task. The algorithm includes handling hundreds of millions of healthy examples, seamless transitions between penalty functions and grouping examples.

1. Dori Peleg is a director and technical fellow at Medtronic, Yokneam site (formerly Given Imaging). He leads the Imaging & Analytics department which is focused on developing algorithms in the machine learning, computer vision and image processing fields.
   Dori Peleg has completed his PhD at the Technion under the supervision of Prof. Ron Meir in the department of electrical engineering. His primary research interests include machine learning techniques for object recognition in images, image classification and optimization.
   

Abstract:

Mammography breast cancer screening over the past three decades has doubled the rate of early breast cancer detection and is credited with decreasing the mortality from breast cancer. Expert screening with current computer-aided detection (CAD) tools has sensitivity estimated between 68%-87% and specificity of 75% 91%. We present a mammography detection algorithm based upon machine learning techniques utilizing both public and proprietary training datasets. We used a committee of diversely filtered shallow Theano-based convolutional neural networks, pre- trained on ImageNet data and fine tuned using False-Color Enhanced Contrast-Limited Adaptive Histogram Equalized mammographic images. Our results yield accuracy of over 98%. 

Abstract

Breathing monitors have become the all-important cornerstone of a wide variety of commercial and personal safety applications, ranging from elderly care to baby monitoring.  Many such monitors exist in the market, some, with vital signs monitoring capabilities, but none remote. We presents a simple, yet efficient, real time method of extracting the subject's breathing sinus rhythm. Points of interest are detected on the subject's body, and the corresponding optical flow is estimated and tracked using the well-known Lucas-Kanade algorithm on a frame by frame basis. A generalized likelihood ratio test is then utilized on each of the many interest points to detect which is moving in harmonic fashion. Finally, a spectral estimation algorithm based on Pisarenko harmonic decomposition tracks the harmonic frequency in real time, and a fusion maximum likelihood algorithm optimally estimates the breathing rate using all points considered. The results show a maximal error of $1$ BPM between the true breathing rate and the algorithm's calculated rate, based on experiments on two babies and three adults.

Nir is a researcher with Essence Group CTO’s office, responsible for bringing to life future technologies for the company. Nir’s research interests are statistical signal processing, namely, estimation, detection theory and machine learning applied to problems in computer vision and radar remote sensing. Nir is also a Ph.D. student in Ben-Gurion University, researching the problem of radar remote sensing, classification and recognition of miniature drones.