A Multimodal Constellation Model for Object Category Recognition

Object category recognition in various appearances is one of the most challenging task in the object recognition research fields. The major approach to solve the task is using the Bag of Features (BoF). The constellation model is another approach that has the following advantages: (a) Adding and changing the candidate categories is easy; (b) Its description accuracy is higher than BoF; (c) Position and scale information, which are ignored by BoF, can be used effectively. On the other hand, this model has two weak points: (1) It is essentially an unimodal model that is unsuitable for categories with many types of appearances. (2) The probability function that represents the constellation model takes a long time to calculate. In this paper we propose a "Multimodal Constellation Model" to solve the two weak points of the constellation model. Experimental results showed the effectivity of the proposed model by comparison to methods using BoF.