A general boosting-based framework for active object recognition

Abstract

We propose a novel general framework with a boosting algorithm to achieve active object classification by view selection. The proposed framework actively decides the next best view for the recognition task. It evaluates different information sources for top hypotheses, generates a voting matrix for candidate views and the view selection is achieved by picking up the one with the maximum votes. Three different sources-similarity based on Implicit Shape Model, prior for model, and prior for views - are presented in the paper. Moreover, we convert view selection itself into a classification problem, and propose a boosting algorithm that is able to combine the previous sources. Experiments show that our algorithm produces a better strategy compared to the other baseline methods.