Abstract

Contextual modeling is a critical issue in scene understand- ing. Ob ject detection accuracy can be improved by exploiting tendencies that are common among ob ject configurations. However, conventional contextual models only exploit the tendencies of normal ob jects; abnor- mal objects that do not follow the same tendencies are hard to detect through contextual model. This paper proposes a novel generative model that detects abnormal ob jects by meeting four proposed criteria of suc- cess. This model generates normal as well as abnormal ob jects, each following their respective tendencies. Moreover, this generation is con- trolled by a latent scene variable. All latent variables of the proposed model are predicted through optimization via population-based Markov Chain Monte Carlo, which has a relatively short convergence time. We present a new abnormal dataset classified into three categories to thor- oughly measure the accuracy of the proposed model for each category; the results demonstrate the superiority of our proposed approach over existing methods.