| 概要： |
Developing computer vision algorithms to interpret scenes of human activity involves a number of related tasks including human detection, tracking, and action recognition. These tasks are intertwined, information from one can provide assist in solving others. In this talk we will describe discriminative latent variable models to address these tasks together, focusing on the latent SVM / max-margin hidden conditional random field. We will present methods for jointly recognizing actions and spatio-temporally localizing them in videos. Models for human-human and human-object interactions will be presented. We will present methods for group activity recognition, with holistic analysis of entire scenes of people interacting and taking different social roles.
|