Indexing of CNN Features for Large Scale Image Search

Convolutional neural network (CNN) features which represent images with global and high-dimensional vectors have shown highly discriminative capability in image search. Although CNN features are more compact than local descriptors, they still cannot efficiently deal with the large-scale image search issue due to its non-negligible computational and storage cost. In this paper, we propose a simple but effective image indexing framework to decrease the computational and storage cost of CNN features. The proposed framework adapts Bag-of-Words model and inverted table to global feature indexing. To this end, two strategies, which are based on the semantic information inside CNN features, are proposed to convert a global vector to one or several discrete words. In addition, a number of strategies for compensating quantization error are fully investigated under the indexing framework. Extensive experimental results on three public benchmarks show the superiority of our framework.