Mixup Augmentation for Deep Hashing
Weida Cao, Futao Liu, Xiaomin Yin, Mingrui Chen, Weizhi Lu, Ran Song · 2021
Deep hashing methods has gained growing popularity in approximate nearest neighbour search for large-scale image retrieval, but exhibit undesirable behaviors such as sensitivity to adversarial examples. In the paper, inspired by the success of mixup-based data augmentation in adversarial training, we for the first time apply this technique to deep hash codes-based image retrieval, and evaluate its performance on six kinds of typical deep hashing methods. Experiments show that the mixup augmentation indeed could provide stable performance gains, ranging from 0.2% to 1.3%.