Towards Large-scale Image Retrieval with a Disk-only Index

Facing ever-growing image databases, the focus of research in content-based image retrieval, where a query image is used to search for those images in a large database that show the same object or scene, has shifted in the last decade. Instead of using local features such as SIFT together with quantization and inverted file indexing schemes, models working with global features and exhaustive search have been proposed to encounter limited main memory and increasing query times. This, however, impairs the capability to find small objects in images with cluttered background. In this paper, we argue, that it is worth reconsidering image retrieval with local features because since then, two crucial ingredients became available: large solid-state disks providing dramatically shorter access times, and more discriminative models enhancing the local features, for example, by encoding their spatial neighborhood using features from convolutional neural networks resulting in way fewer random read memory accesses. We show that properly combining both insights renders it possible to keep the index of the database images on the disk rather than in the main memory which allows even larger databases on today's hardware. As proof of concept we support our arguments with experiments on established public datasets for large-scale image retrieval.