Rating is available when the video has been rented.
This feature is not available right now. Please try again later.
Published on Jun 24, 2012
CVPR 2012 Spotlight Video. Sameh Khamis, presenting joint work with Vlad Morariu and Larry Davis.
Transcript: This is Sameh Khamis, and this is joint work with Vlad Morariu and Larry Davis at the university of Maryland. We propose a model to couple tracking and action recognition. Bounding boxes of people with similar pose will likely generate similar feature responses, making it hard to discriminate between different actions. So while tracking can benefit action recognition, we also observe that action recognition can benefit tracking. To solve this problem in the context of multi-person action recognition, we propose a network flow-based model, where the minimum cost flow decodes to a unique assignment of actions and identities. The corresponding cost function prefers appearance consistency, which means a person tends to look similar across frames, and action compatibility, which means actions transition naturally across frames, so waiting to cross the street is likely to be followed by crossing rather than dancing. Our model is tractable and efficient, and we report state-of-the-art results on two publicly available datasets.