Approximate Nearest Neighbor Search by Residual Vector Quantization

Abstract

A recently proposed product quantization method is efficient for large scale approximate nearest neighbor search, however, its performance on unstructured vectors is limited. This paper introduces residual vector quantization based approaches that are appropriate for unstructured vectors. Database vectors are quantized by residual vector quantizer. The reproductions are represented by short codes composed of their quantization indices. Euclidean distance between query vector and database vector is approximated by asymmetric distance, i.e., the distance between the query vector and the reproduction of the database vector. An efficient exhaustive search approach is proposed by fast computing the asymmetric distance. A straight forward non-exhaustive search approach is proposed for large scale search. Our approaches are compared to two state-of-the-art methods, spectral hashing and product quantization, on both structured and unstructured datasets. Results show that our approaches obtain the best results in terms of the trade-off between search quality and memory usage.

Compact Hash Codes for Efficient Visual Descriptors Retrieval in Large Scale Databases

“…6) and GIST1M (Fig. 7) datasets against several methods presented in Section II: Product Quantization (ADC and IVFADC) [1], PQ-RO [19], PQ-RR [19], Cartesian k-means [15], OPQ-P [16], [51], OPQ-NP [16], [51], LOPQ [18], a non-exhaustive adaptation of OPQ [16], called I-OPQ [18], RVQ [52] , RVQ-P [19] and RVQ-NP [19].…”

“…RVQ [52] approximates the quantization error by another quantizer instead of discarding it. In this method several stagequantizers, each one with its corresponding stage-codebook, are connected sequentially.…”

PharmaPack: Mobile fine-grained recognition of pharma packages

“…neural codes [5], or aggregation of low-or-mid-dimensional local descriptors such as SIFT [6], SURF [7], aKaZe [8], etc. using feature aggregation such as Fisher vectors [9], VLAD [10], residual vectors [11], triangulated embedding [12], etc., that produces a resulting vector of defined length D. The resulting high-dimensional descriptors are collected in a database consisting of N enrolled feature vectors and the identification system should produce a list L(q) of indices of enrolled features f (i) ∈ R D , 1 ≤ i ≤ N closest to the probe feature vector q ∈ R D . Additionally, local descriptors can be stored together with their coordinates within the image.…”

Retrieval of Binary Features in Image Databases: A Study

“…Quantization is usually performed with variants of kMeans clustering [25]. With larger vocabularies, however, quantization based on kMeans clustering becomes more and more costly, as assigning a value to its representant is linear in k. To reduce the high complexity of assigning feature vectors to cluster centers during query evaluation, a variety of non-optimal but fast quantization techniques such as Hierarchical kMeans [17,24], approximate kMeans [21], Product Quantization [11] and Residual Vector Quantization [5] have been proposed. After employing quantization for BoVW-based retrieval, it has also been investigated in the context of kNN search [11], which has since then received quite some attention.…”