Abstract:  Object detection and recognition is a hot topic in computer vision.Traditional methods use only local features for detection. Recently, some research results show that the detecting performance could be improving by using topic features. Some researchers employed topic models which is originally used for text analysis to extract topic features from images for object recognition and detection. However, visual features should be quantized into virtual words and information of class label should be ignored while using traditional topic models such as probabilistic latent somatic analysis(PLSA)，latent Dirichlet allocation(LISA)and so on, In order to utilize continuous local features and information of class label in one model，we propose a new graphical model named latent Dirichlct classification(LDC)，which is inspired by LDA model.The proposed model has three more variables than LDA in the graphical structure; x(local features)，‘(class label) and v(parameter). In the proposed model，we consider class label of each image block is determined by both of its local features and topic features based on original
LDA model. Parameter v is a set of classifiers trained for combining these two features. Similar with the inference process of LDA model，we use variational inference to solve our model. As a result of continuous local features，information of class label and topic features arc all token into consideration reasonably, LDC can be used in object detection directly and efficiently, In the end of this paper, we test the availability of LDC on two datasets. Experimental results show that our proposed model improve the performance of object detection efficiently.