Analysis of the Effectiveness-Efficiency Dependance for Image Retrieval

Abstract

Similarity search in image database is commonly implemented as nearest-neighbor search in a feature space of the images. For that purpose, a large number of different features as well as different search algorithms have been proposed in literature. While the efficiency aspect of similarity search has attracted a great interest in the past few years, the effectiveness of the search was often neglected. In this work, however, we argue that these two measures interplay with each other. The longer the feature representation is, the better the quality of the retrieval gets, but the larger the execution costs become. In other words, an improvement in effectiveness leads to a deterioration of performance and vice versa. The aim of this work is to explicitly take both measures into account to optimize the retrieval both form a quality perspective and a performance perspective. To this end, we define a benchmark including a measure for the efficiency and the effectiveness of a feature. Then one can compare different features or feature combinations using simple two-dimensional plots. Based on the quality and performance constraints of a user, the search engine can easily determine the optimal feature or feature combination. Finally, we have applied our benchmark to a large number of different feature types to compare their effectiveness-efficiency relationship.