Voronoi-based compact image descriptors: Efficient Region-of-Interest retrieval with VLAD and deep-learning-based descriptors

摘要

We investigate the problem of image retrieval based on visual queries whenthe latter comprise arbitrary regions-of-interest (ROI) rather than entireimages. Our proposal is a compact image descriptor that combines thestate-of-the-art in content-based descriptor extraction with a multi-level,Voronoi-based spatial partitioning of each dataset image. The proposedmulti-level Voronoi-based encoding uses a spatial hierarchical K-means overinterest-point locations, and computes a content-based descriptor over eachcell. In order to reduce the matching complexity with minimal or no sacrificein retrieval performance: (i) we utilize the tree structure of the spatialhierarchical K-means to perform a top-to-bottom pruning for local similaritymaxima; (ii) we propose a new image similarity score that combines relevantinformation from all partition levels into a single measure for similarity;(iii) we combine our proposal with a novel and efficient approach for optimalbit allocation within quantized descriptor representations. By deriving both aVoronoi-based VLAD descriptor (termed as Fast-VVLAD) and a Voronoi-based deepconvolutional neural network (CNN) descriptor (termed as Fast-VDCNN), wedemonstrate that our Voronoi-based framework is agnostic to the descriptorbasis, and can easily be slotted into existing frameworks. Via a range of ROIqueries in two standard datasets, it is shown that the Voronoi-baseddescriptors achieve comparable or higher mean Average Precision againstconventional grid-based spatial search, while offering more than two-foldreduction in complexity. Finally, beyond ROI queries, we show that Voronoipartitioning improves the geometric invariance of compact CNN descriptors,thereby resulting in competitive performance to the current state-of-the-art onwhole image retrieval.