Object recognition using multiple instance learning with unclear object teaching

Abstract

We propose an object recognition method for service robots under the constraint of uncertain object teaching by humans. In previous object recognition methods, the training phase required a large number of prepared images and also required the training data to not have a complex background. However, for robots to perform daily tasks, they should be able to recognize objects despite unclear object teaching by humans. In order to mitigate the effect of features in the background on object recognition, our proposed method classifies local features based on saliency from video images. In this paper, we demonstrate the efficacy of the proposed method in recognizing target objects despite unclear teaching by the user.