A study on parameter tuning for optimal indexing on large scale datasets

Abstract

Fast matching is a crucial task in many computer vision applications due to its computationally intensive overhead, especially for high feature spaces. Promising techniques to address this problem have been investigated in the literature such as product quantization, hierarchical clustering decomposition, etc. In these approaches, a distance metric must be learned to support the re-ranking step that helps filter out the best candidates. Nonetheless, computing the distances is a much intensively computational task and is often done during the online search phase. As a result, this process degrades the search performance. In this work, we conduct a study on parameter tuning to make efficient the computation of distances. Different searching strategies are also investigated to justify the impact of coding quality on search performance. Experiments have been conducted in a standard product quantization framework and showed interesting results in terms of both coding quality and search efficiency.