Effective training of convolutional networks using noisy Web images

Abstract

Deep convolutional networks have recently shown very interesting performance in a variety of computer vision tasks. Besides network architecture optimization, a key contribution to their success is the availability of training data. Network training is usually done with manually validated data but this approach has a significant cost and poses a scalability problem. Here we introduce an innovative pipeline that combines weakly-supervised image reranking methods and network fine-tuning to effectively train convolutional networks from noisy Web collections. We evaluate the proposed training method versus the conventional supervised training on cross-domain classification tasks. Results show that our method outperforms the conventional method in all of the three datasets. Our findings open opportunities for researchers and practitioners to use convolutional networks with inexpensive training cost.