Object categorization using context from multi-spatial levels

To categorize objects in the real-world scene images, a method is proposed by exploiting multi-spatial extent context. Firstly, a soft decision-based sampling mechanism is utilized in the local image patch sampling process, by which, mixed information in the scene can be separated in an effective and robust way. Then, by using the soft decision-based sampling mechanism and the statistical representation methods, the statistical feature for each spatial extent context can be computed. Finally, a logistic regression classification method is adopted to integrate multiple spatial extent context information and make the final decisions. The experiments show that, the proposed method can better model the objects in the real world scenes, and thus apparently improves the object categorization performance.