Accelerated Nearest Neighbor Search with Quick ADC

Abstract

E cient Nearest Neighbor (NN) search in high-dimensional spaces is a foundation of many multimedia retrieval systems. Because it o ers low responses times, Product Quantization (PQ) is a popular solution. PQ compresses high-dimensional vectors into short codes using several sub-quantizers, which enables in-RAM storage of large databases. This allows fast answers to NN queries, without accessing the SSD or HDD. The key feature of PQ is that it can compute distances between short codes and high-dimensional vectors using cache-resident lookup tables. The e ciency of this technique, named Asymmetric Distance Computation (ADC), remains limited because it performs many cache accesses.In this paper, we introduce Quick ADC, a novel technique that achieves a 3 to 6 times speedup over ADC by exploiting Single Instruction Multiple Data (SIMD) units available in current CPUs. Efciently exploiting SIMD requires algorithmic changes to the ADC procedure. Namely, Quick ADC relies on two key modi cations of ADC: (i) the use 4-bit sub-quantizers instead of the standard 8-bit sub-quantizers and (ii) the quantization of oating-point distances. This allows Quick ADC to exceed the performance of state-of-theart systems, e.g., it achieves a Recall@100 of 0.94 in 3.4 ms on 1 billion SIFT descriptors (128-bit codes).

Quicker ADC : Unlocking the Hidden Potential of Product Quantization With SIMD

“…In this paper, we presented Quicker ADC, a novel distance computation method for product-quantization-based ANN search. Quicker ADC improves over previous proposition [2] by (i) supporting additional quantizers (e.g., m×{6, 6, 4} , m×{8, 8} , m×{8} ) and (ii) having an improved implementation integrated into FAISS and compatible with various indexes (IMI, IVF HNSW). Through an extensive evaluation, we have shown that Quicker ADC outperforms schemes based on PQ or polysemous codes for both exhaustive and non-exhaustive (i.e., index-based) search, and that they combine well with the latest indexes such as HNSW-based IVF [27].…”

“…PQ Fast Scan [1] pioneered the use of SIMD for ADC distance evaluation. It inspired later work [2], [11], [32] that are more suitable for indexed databases. Quicker ADC is a generalization of Quick ADC [2] supporting additional shuffles and an improved distance quantization scheme.…”

“…This loss of accuracy comes mainly from the reduced precision of 4-bit subquantizers when compared to 8-bit subquantizers and to a smaller extent from the use of quantized distances. Yet, [2], [11], [32] remain limited to m×4 as they follow the initial approach of [1] that use pshufb.…”

Distributed and Consistent Multi-Image Feature Matching via QuickMatch

“…The knearest neighbor problem has been extensively explored in settings such as friend suggestion on Facebook, image classification, and recommendation systems. For this reason, many parallelized approaches have been introduced by Johnson (2019), Andre (2017), andGieseke (2014). In a feature matching setting, a CUDA based distributed approach to the BF algorithm has been proposed that is approximately 100x faster than its centralized counterpart.…”

Distributed and consistent multi-image feature matching via QuickMatch

“…The k -nearest neighbor problem has been extensively explored in settings such as friend suggestion on Facebook, image classification, and recommendation systems. For this reason, many parallelized approaches have been introduced by Johnson et al (2019), Andre et al (2017), Gieseke et al (2014), and others. In a feature matching setting, a CUDA-based distributed approach to the BF algorithm has been proposed that is approximately 100× faster than its centralized counterpart.…”