Combining Self-Aggregation Vector and Indicator Matrix for Improved Object Retrieval

Deep features extracted by convolutional neural networks have demonstrated state-of-the-art performance on image classification and retrieval tasks. Firstly, this work presents a simple and efficient method to select the region of interest based on semantics of features. Then, this work spatially weight deep convolutional features based on analyzing the relations between the self-aggregation vector and the indicator matrix. Based on the two strategies mentioned above, we product a compact global image representation of all images and show the proposed approach outperforms previous strategies based on pre-trained networks on a couple of benchmark datasets, i.e., Oxford dataset and Paris dataset.