Graphical models for context-aware analysis of continuous videos

Abstract

In this paper, we show how graphical models can used for the localization and recognition of activities in continuous videos. The model consists of an action layer and a hidden activity layer. The action layer is modeled as a linear-chain conditional random field (CRF) with the activity labels of action segments as the model variables. Hidden activity variables are then introduced to smooth out the activity labels of action segments and thus generating semantically meaningful activities. With a task-oriented discriminative approach, the learning problem is formulated as a latent Structural Support Vector Machine (SSVM). We show promising results on the UCLA Office Dataset that demonstrate the effectiveness of the proposed framework.