Memory-efficient NBNN image classification

Abstract

NBNN is a simple image classifier based on identifying nearest neighbors. NBNN uses original image descriptors (e.g., SIFTs) without vector quantization for preserving the discriminative power of descriptors and has a powerful generalization characteristic. However, it has a distinct disadvantage

its memory requirement can be prohibitively high as we have a large amount of data. We identify this problem of NBNN techniques and we apply a binary code embedding technique, i.e., spherical hashing, to encode data compactly without a significant loss of classification accuracy. We also propose to use an inverted index to identify nearest neighbors among those binarized image descriptors. To demonstrate benefits of our method, we apply our method to two of existing NBNN techniques with a image dataset. By using 64~bit lengths, we are able to observe 16 times memory reduction with a higher performance without a significant loss of classification accuracy. This result is thanks to our compact encoding of image descriptors without losing much information of original image descriptors.