Active view selection for object and pose recognition

Abstract

In this paper we present an algorithm for multi-view object and pose recognition. In contrast to the existing work that focuses on modeling the object using the images only; we exploit the information on the image sequences and their relative 3D positions, because under many circumstances the movements between multi-views are accessible and can be controlled by the users. Thus we can calculate the next optimal place to take a picture based on previous behaviors, and perform the object/pose recognition based on these obtained images. The proposed method uses HOG (Histograms of Oriented Gradient) and SVM (Support Vector Machine) as the basic object/pose classifier. To learn the optimal action, this algorithm makes use of a boosting method to find the best sequence across the multi-views. Then it exploits the relation between the different view points using the Adaboost algorithm. The experiment shows that the learned sequence improves recognition performance in early steps compared to a randomly selected sequence, and the proposed algorithm can achieve a better recognition accuracy than the baseline method.