A new web-supervised method for image dataset constructions

Abstract

The goal of this work is to automatically collect a large number of highly relevant natural images from Internet for given queries. A novel automatic image dataset construction framework is proposed by employing multiple query expansions. In specific, the given queries are first expanded by searching in the Google Books Ngrams Corpora to obtain a richer semantic descriptions, from which the visually non-salient and less relevant expansions are then filtered. After retrieving images from the Internet with filtered expansions, we further filter noisy images by clustering and progressively Convolutional Neural Networks (CNN) based methods. To evaluate the performance of our proposed method for image dataset construction, we build an image dataset with 10 categories. We then run object detections on our image dataset with three other image datasets which were constructed by weak supervised, web supervised and full supervised learning, the experimental results indicated the effectiveness of our method is superior to weak supervised and web supervised state-of-the-art methods. In addition, we do a cross-dataset classification to evaluate the performance of our dataset with two publically available manual labelled dataset STL-10 and CIFAR-10.