Part-based Weighting Aggregation of Deep Convolutional Features for Image Retrieval

Several recent works have shown that part-based image representation provides state-of-the-art performance for fine-grained categorization. Moreover, it has also been shown that image global representation generated by aggregating deep convolutional features provides excellent performance for image retrieval. In this paper we propose a novel aggregation method, which utilizes the information of retrieval object parts. The proposed part-based weighting aggregation (PWA) method utilizes the normalized feature maps as part detectors to weight and aggregate the convolutional features. The part detectors which are selected by the unsupervised method highlight the discriminative parts of objects and effectively suppress the noise of background. We experiment on five public standard datasets for image retrieval. Our unsupervised PWA method outperforms the state-of-the-art approaches based on pre-trained networks and achieves comparable accuracy with the fine-tuned methods. It is worth noting that our unsupervised method is very suitable and effective for the situation where the annotated training dataset is difficult to collect.