Automatic online labeling images via co-active-learning

The well-built dataset is a pre-requisite for computer vision research. However, the process of collecting and labeling the images is laborious and monotonous. In this paper, we aim to automatic labeling and collecting the images for the visual object category. We propose an active learning approach in a co-training way. Our approaches starts from a small dataset with ground truth labels, and iteratively labels a larger set of unlabeled samples using the two irrelative classifiers and augment the labeled dataset, and update the learning model simultaneously. There are two advantages of our approach, one is to avoid drifting from the object category, and the other is to sequentially update the learning model with the increasing of the unlabeled samples. The experiment results demonstrate that our approach is effective and is superior to the self-training methods.