Co-weighting semantic convolutional features for object retrieval

Deep feature aggregation, which refers to aggregating a set of local convolutional features into a global image-level vector, has attracted increasing attention in object instance retrieval. In this manuscript, we propose an unsupervised framework that aggregates feature maps by an adaptive selection and two weighting strategies. Particularly, the selection process finds the foreground contour by explaining the semantic structure implicated in the feature maps, while two weighting process including an adaptive Gaussian filter that highlights semantic features and an element-value sensitive channel vector that activates feature channels corresponding to sparse yet distinctive image patterns. Experimental results on benchmark image retrieval datasets validate that the selection and two weighting schemes are complementary in improving the discriminative power of image vectors. With the same experimental settings, the proposed approach outperforms state-of-the-art aggregation approaches by a considerable margin.