Weighted two-step aggregated VLAD for image retrieval

The vector of locally aggregated descriptor (VLAD) has been demonstrated to be efficient and effective in image retrieval and classification tasks. Due to the small-size codebook adopted by the method, the feature space division is coarse and the discriminative power is limited. Toward a discriminative and compact image representation for visual search, we develop a novel aggregating method to build VLAD, called two-step aggregated VLAD. Firstly, we propose the bidirectional quantization from both views of descriptors and visual words, for getting finer division of feature space. Secondly, we impose the probabilistic inverse document frequency to weight the local descriptors, for highlighting the discriminative ones. Experimental results on extensive datasets show that our method yields significant improvement and is competitive with the state-of-the-art methods.