Mixture models for object detection

Abstract

In this paper, we propose an approach based on mixture of multiple components and mid-level part models for object detection in natural scenes. It is difficult to represent an object category with a monolithic model as the intravariance in the category. To solve this, we use multi-component models and part models to describe the global variation and local deformation respectively. We obtain multi-components by clustering to form visual similar object group and training discriminant model for each one. The mid-level part models are learned automatically. We apply max-pooling to generate the feature vector using all part models and then train the SVM classifier based on these feature vectors. When detecting in image, we first achieve object candidates using multi-component models, and then the performance is refined by using part models and SVM classifier. Experiments on standard benchmarks demonstrate this coarse-to-fine detection system performs competitively.