Event Location: NSH 1507
Bio: Yair Movshovitz-Attias is a PhD student in the Computer Science Department, advised by Prof. Yaser Sheikh and Prof. Takeo Kanade. His research interests include object detection, image classification and fine-grained recognition. Yair has an MSc in Computer Science and a BSc in Computer Engineering from the Hebrew University of Jerusalem. He has worked as a software engineer, a computer technician and even a video game tester.

Abstract: Cars have a central role in our lives. Together with cell phones and computers they are, arguably, the most socially significant objects we interact with. Given their significance they provide a set of opportunities that are unlike any other object studied in Computer Vision.

In this talk I will focus on one important benefit of using cars as a test case for computer vision research – The abundance of real world knowledge we have about cars. First, I will show how to leverage high quality 3D CAD models of cars for accurate object viewpoint estimation. Our approach estimates a sparse basis of correlation filter detectors designed to directly maximize detection rate, while reducing the number of detectors. This method provides the benefits of exemplar-based approaches, while significantly reducing the computational complexity. We present state-of-the-art results on three publicly available datasets and introduce a new dataset for continuous car pose estimation with over 3000 street-scene images. Secondly, I will discuss work in progress on a method that incorporates information collected from Internet Knowledge Bases to improve fine-grained classification. We extract various car attributes and model the structure encoded by the information using a Bayes Net graphical model that propagates information on the attributes. Lastly, I will describe a mobile app we have built to crowd source the collection of fine-grained labels.