Automatic Database Creation and Object’s Model Learning

Abstract

This paper proposes a new framework to automatically generate visual object database meanwhile efficiently learn the object’s model. The system is of important need for the problems of object detection and recognition. Our main idea is to acquire the huge amount of video data actively, and seeks out opportunities to autonomously exploit information from object samples. We employ autonomous learning approach based on online boosting technique, which allows to combine an object detector trained on a single initialized input image with tracking to extract object samples for learning. The autonomous learning process with interactive learning strategy allows to adaptively improve the learning object model while generating informative samples. Our method allows to generate thousands of object samples within hours from large video databases or from live camera, thus saving time and labor’s efforts. We will show that the proposed method can extracts well-localized, diverse appearances of object examples from video sequence through only one initialized input sample, and builds robust object model. In addition to requiring very little human intervention, a significant benefit of this method is that it does not require pre-training. In the experiments, the approach is evaluated in detail for creating data sets and learning for the problems of human hand gesture recognition and face detection. In addition, to show the generality, results for different objects are also presented.