CVPR Daily - Wednesday

Amit Kumar is a graduate student at the University of Maryland, working with Professor Rama Chellappa. He spoke to us during his poster session yesterday. Amit’s work is about detecting key points from faces. Given an unconstrained image, which can be in any arbitrary pose, it can detect key points and those key points can be used for many purposes including aligning the image for face verification . He explains that given an image, this work aims to understand how the pose of a face impacts the key points and how to model the relationship between different key points. Our faces have an inherent structure – two eyes on the top and a nose and a mouth – but how do we model that structure in a network? The challenge in doing this, he says, are the unconstrained faces, due to low-lighting and low-resolution images, and images where the person is not even visible. Amit explains what algorithmic techniques are used: “ Basically maths, as in probability. How I condition my key points on 3D pose, that’s one mathematical fundamental. Then the deep learning convolutional neural networks. ” He is doing this work right now for 2D images, but in terms of next steps, Amit would like to extend it to generate 3D models of faces . He says he can use these key points to constrain and get a good 3D model, instead of some of the strange looking 3D models that are available now. “Those key points can be used for many purposes, including aligning the image for face verification. ” 10 Amit Kumar Wednesday