An improved ANN search algorithm for visual search applications

Approximate nearest neighbor search is a kind of significant algorithm to ensure the accuracy and speed for visual search system. In this paper, we ameliorate the search algorithm following the framework of product quantization. Product quantization can generate an exponentially large codebook by a product quantizer and then achieve rapid search with the asymmetric distance computation or symmetric distance computation, while it will still produce a larger distortion in some cases when calculating the approximate distance. Therefore, we design the hierarchical residual product quantization which simultaneously quantifies the input and residual space and meanwhile we extend the asymmetric distance computation to handle this quantization method which is still very efficient to estimate the approximate distance. We have tested our method on several datasets, and the experiment shows that our method consistently improves the accuracy against the-state-of-the-art methods.