Skip to content. Skip to navigation
CIM Menus
 

Using Unlabelled 3D Motion Examples for Human Activity Understanding


Ankur Gupta
University of British Columbia

March 11, 2016 at  10:30 AM
McConnell Engineering Room 320

Abstract:

Recognizing activities and detecting human pose in monocular videos are central problems in computer vision with wide-ranging applications in robotics, sports analysis, and healthcare. Obtaining annotated data to learn from videos in a supervised manner is tedious, time-consuming, and not scalable to a large number of human activities. To deal with these issues, we propose an unsupervised, data-driven approach which only relies on unlabelled 3d motion examples (mocap sequences).
The first part of the talk deals with adding view-invariance to the standard action recognition task, i.e., identifying the class of activity given a short video sequence. We build a view-invariant representation of human motion by generating synthetic features using mocap data. Next, we focus on the problem of mocap retrieval using realistic videos as queries. Retrieving similar mocap sequences from a large database can be useful for character animation, 3d pose tracking, and cross-view action recognition. We present a novel and scalable approach to the problem, utilizing a discriminatively-trained 2d pose detector as well as optical flow-based motion features.

Bio:

Ankur Gupta is a Ph.D. candidate at the University of British Columbia. He is co-supervised by Prof. Jim Little and Prof. Bob Woodham. Ankur completed his bachelor's degree in Electrical Engineering at the Indian Institute of Technology, Kanpur, India. His main research interests are human activity understanding and 3d pose estimation.