Image Parsing with Multilevel Representations

I will describe recent and ongoing work on parsing visual scenes using rich, multi-scale representations. Such representations emerge from the structure of deep convolutional neural networks trained on large sets of natural images. I will focus on two tasks: semantic segmentation, in which image pixels are labeled according to semantic categorization of regions they belong to; and automatic image colorization, where the goal is to predict meaningful colors for every pixel of a grayscale input image. I will also describe a connection between the two tasks, and time permitting, briefly review some other ongoing projects related to representation learning.

Greg Shakhnarovich is an Associate Professor at the Toyota Technological Institute at Chicago, working on computer vision and machine learning. He also holds a part-time faculty appointment in the Computer Science Department, University of Chicago. Greg came to TTIC in 2008 from Brown, where he was a postdoctoral researcher, after obtaining a PhD in Computer Science from MIT in 2005.