Local Features Augmenting for Better Image Retrieval

Recently, a lot of works have shown the advantages of utilizing the deep descriptors, obtained from the features of the last convolution layer in CNNs, on image retrieval. In this paper, we focus on augmenting and fusing CNN features for the image retrieval task. We first investigate the effects of network rotation, and then propose two models for deep feature augmenting: single model augmenting and multiple model augmenting. For the single model augmenting, we expand the model by rotating and flipping the single network. While for the multiple model, we expand filters by connecting the different networks together. As to the fusion methods, we evaluate concatenation, average and max pooling. We conduct a thorough evaluation of the above models and fusion approaches, and show the state of the art performance of our proposed approach.