Active Learning the Model for Indoor Scene Classification

Abstract

Scene classification problem have became a popular topic in computer vision and robotics society. When faced with large amount of unlabeled data, manual annotation is sometimes difficult and time-consuming. Active learning is one of the techniques that can automatically identify the informative instances to be selected for labeling, which can greatly reduce human effort in inducing classifiers. In this work, we propose a novel active learning approach based on a Relevance Vector Machine (RVM) framework, which is then applied to the tasks of indoor scene classification. The query criterion of this method takes entropy-based uncertainty and representative ness of instances into account simultaneously, which can be easy computed under RVM modeling. Our results on a four-class indoor scene dataset reveal the efficacy of the proposed framework.