Gathering & Poster Session
Prof. David Malah, Head of SIPL
Improving Training Efficiency in Deep Learning
Prof. Daniel Soudry
I will present several empirical and theoretical results related to improving training efficiency in deep networks, based on the following works:
- Elad Hoffer, Itay Hubara, and Daniel Soudry, “Fix your classifier: the marginal value of training the last weight layer.” ICLR 2018.
- Elad Hoffer, Ron Banner, Itay Golan, and Daniel Soudry, “Norm matters: efficient and accurate normalization schemes in deep networks.” arXiv 2018.
- Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry, “Scalable methods for 8-bit training of neural networks.” arXiv 2018.
Wilk Family Awards and Outstanding Supervisor Awards Ceremony
Unsynchronized Acoustic Indoor Positioning
Finalist in the Kasher undergraduate project contest in the Faculty of Electrical Engineering
Guy Feferman, Michal Blatt
Supervisor: Alon Eilam
In cooperation with:
ICASSP 2018 Demo
Break & Poster Session
Review of Teaching Activity in SIPL
Detection and Localization of Cumulonimbus Clouds in Satellite Images
Wilk family award winner
Etai Wagner, Ron Dorfman
Supervisor: Almog Lahav
In cooperation with:
Submitted to ICSEE 2018
Local-to-Global Point Cloud Registration using a Viewpoint Dictionary
David Avidar, M.Sc. student
Advisors: Prof. David Malah, Dr. Meir Bar-Zohar
Partly funded by the OMEK consortium
Presented at ICCV 2017
Local-to global point cloud registration is a challenging task due to the substantial differences between these two types of data, and the different techniques used to acquire them. Global clouds cover large-scale environments and are usually acquired aerially (e.g., using Airborne Laser Scanning – ALS), and local clouds are often acquired from ground level at a much smaller range (e.g., using Terrestrial Laser Scanning – TLS). As a result of these differences, existing point cloud registration approaches, such as keypoint-based registration, tend to fail.
We propose a novel registration method based on converting the global cloud into a viewpoint-based dictionary. We associate each viewpoint with a panoramic range-image, capturing the geometry of the visible environment. Then, plausible local-to-global transformations can be found via a dictionary search. We show efficient dictionary search can be done using phase-correlation between panoramic range-images.
We demonstrate that the proposed viewpoint-dictionary-based registration method achieves better performance than state-of-the-art, keypoint-based methods (e.g., FPFH, RoPS), even without any GPS measurements. For the evaluation, we used a challenging dataset of 108 TLS local clouds and an ALS large-scale global cloud, in a 1km2 urban environment.
The Perception-Distortion Tradeoff
Yochai Blau, Ph.D. student
Advisor: Prof. Tomer Michaeli
Presented at CVPR 2018
Image restoration algorithms are typically evaluated by some distortion measure (e.g. PSNR, SSIM, IFC, VIF) or by human opinion scores that quantify perceived perceptual quality. In this work, we prove mathematically that distortion and perceptual quality are at odds with each other. Specifically, we study the optimal probability for correctly discriminating the outputs of an image restoration algorithm from real images. We show that as the mean distortion decreases, this probability must increase (indicating worse perceptual quality). As opposed to the common belief, this result holds true for any distortion measure, and is not only a problem of the PSNR or SSIM criteria. However, as we show experimentally, for some measures it is less severe (e.g. distance between VGG features). We also show that generative-adversarial-nets (GANs) provide a principled way to approach the perception-distortion bound. This constitutes theoretical support to their observed success in low-level vision tasks. Based on our analysis, we propose a new methodology for evaluating image restoration methods, and use it to perform an extensive comparison between recent super-resolution algorithms. Our study reveals which methods are currently closest to the theoretical perception-distortion bound.