Abstract

We propose a new vector encoding scheme (tree quantization) that obtains lossy compact codes for highdimensional vectors via tree-based dynamic programming. Similarly to several previous schemes such as product quantization, these codes correspond to codeword numbers within multiple codebooks. We propose an integer programming-based optimization that jointly recovers the coding tree structure and the codebooks by minimizing the compression error on a training dataset. In the experiments with diverse visual descriptors (SIFT, neural codes, Fisher vectors), tree quantization is shown to combine fast encoding and state-of-the-art accuracy in terms of the compression error, the retrieval performance, and the image classi- fification error.