Binary Constrained Deep Hashing Network for Image Retrieval without Human Intervention.

Learning compact binary codes for image retrieval problem using deep neuralnetworks has attracted increasing attention recently. However, training deephashing networks is challenging due to the binary constraints on the hashcodes, the similarity preserving properties, and the requirement for a vastamount of labelled images. To the best of our knowledge, none of the existingmethods has tackled all of these challenges completely in a unified framework.In this work, we propose a novel end-toend deep hashing approach, which istrained to produce binary codes directly from image pixels without humanintervention. In particular, our main contribution is to propose a novelpairwise loss function, which simultaneously encodes the distances betweenpairs of binary codes, and the binary quantization error. We propose anefficient parameter learning algorithm for this loss function. In addition, toprovide similar/dissimilar images for our pairwise loss function, we exploit 3Dmodels reconstructed from unlabeled images for automatic generation of enormoussimilar/dissimilar pairs. Extensive experiments on three image retrievalbenchmark datasets demonstrate the superior performance of the proposed method.