A team of researchers from the LTI and the University of Technology, Sydney (UTS), recently earned first place in the Action Classification Task category of the 2015 THUMOS Challenge. Held in conjunction with the Conference on Computer Vision and Pattern Recognition, the THUMOS Challenge aims to automatically recognize a large number of human actions from open source videos in a realistic setting. LTI members participating on the CMU-UTS team included Principal Systems Scientist Alex Hauptmann, former post-doc Yi Yang and visiting scholar Zhongwen Xu.
For the Action Classification task challenge, teams were presented with a list of actions that ranged from daily activities, like brushing teeth, to sports actions like driving. They then developed a system that could automatically predict the presence or absence of these actions in a given video, and then associated that prediction with an accuracy score. The team with the highest accuracy score won.
According to Hauptmann, activity recognition is a fundamental building block of understanding what people or animals are doing while observed. Successful automated analysis has strong implications for medical diagnosis and rehabilitation support, surveillance systems, and understanding the content of video in general. This critical understanding has been difficult to achieve, however, due to the variability in where actions take place, how they are performed and how they are observed.
"Activity recognition is important for understanding anything humans do, yet until recently computer systems have been quite bad at it, outside of very constrained circumstances," Hauptmann said. "This work shows a path toward overcoming this inability and opens the door for significant applications in the future."