Bio: Nima Razavi is currently a post-doctoral researcher in Luc van Gool’s lab at ETH Zurich where he finished his PhD in 2012. He holds a M.Sc. in Computer Science from ETH Zurich and a B.Sc. in computer engineering from Sharif University of Technology. His research interests are in Computer Vision, Machine Learning and Neuroscience and in particular the problem of object recognition and scene understanding.

Abstract: In this talk, I will present our recent works on developing a scalable voting-based object detector with a shared vocabulary of parts. Scalability of object detectors with respect to the number of classes and training images is a critically important issue for applications with a large number of classes. Detecting objects with a shared vocabulary of parts is also very attractive as it facilitates better generalization and incremental learning of the detector.

After giving a complexity analysis of detection with the Hough transform, I will derive the necessary criteria for the vocabulary to be scalable and show how to train it by directly maximizing these criteria. When using Hough transform for detecting objects with a shared vocabulary, it is crucial to only allow for consistent parts to form an object hypothesis. I will show how to enforce consistency of parts by augmenting the Hough space with latent variables, and discriminatively learn a Latent Hough Transform using a validation set.