Discriminative models for static human-object interactions

Abstract

We advocate an approach to activity recognition based on modeling contextual interactions between postured human bodies and nearby objects. We focus on the difficult task of recognizing actions from static images and formulate the problem as a latent structured labeling problem. We develop a unified, discriminative model for such context-based action recognition building on recent techniques for learning large-scale discriminative models. The resulting contextual models learned by our system outperform previously published results on a database of sports actions.