Drinking activity analysis from fast food eating video using generative models

Abstract

The drinking activity is a common event in a fast food eating process. In this paper, we present a study on drinking activity analysis from fast food eating video using generative models. We apply three different generative models, namely Conditional Random Field (CRF), Hidden-state Conditional Random Field (HCRF), and Latent-Dynamic Conditional Random Field (LDCRF), to characterize drinking activities in a fast food eating process. The CRF and LDCRF models are applied in the frame and sequence level classification while HCRF model is used on video clip classification. We evaluate the proposed method on a dataset that contains 27 videos from 9 fast food restaurants. Experimental results show that the proposed method can obtain promising classification results for drinking activity labeling and classification, averagely achieving accuracies of 79.80% for frame and 72.81% for subsequence. And the CRF model outperforms LDCRF and HCRF models.