Bootstrapping Image Classification with Sample Evaluation

In this work, we look at the problem of multi-class image classification in a semisupervised learning framework. Given a small set of labeled images, and a much larger set of unlabeled images, we propose a semi-supervised learning method that combines bootstrapping with sample evaluation, to continuously update the learned models for each class. Bootstrapping involves using self-labeled images to re-train the learned models. To overcome the semantic drift that naive bootstrapping is prone to, we use additional sample evaluation methods based on the ideas of co-training and pairwise constraints, to determine whether or not a newly classified instance should be used for re-training. Experimental results show the usefulness of sample evaluation, when used in conjunction with bootstrapping. In particular, our method is able to achieve a 8% improvement in overall accuracy over baseline bootstrapping, on a 15 class subset of the SUN (Scene UNderstanding) dataset.