||Riemenschneider Hayko, Donoser Michael, Bischof Horst
||Computer Vision Winter Workshop (CVWW)
||Content-based image retrieval (CBIR) moves retrieval in large databases from metadata to actual visual content. There are two main paradigms for the type of visual input: query-by-example and query-by-sketch. Most retrieval systems adopt the former and perform a near duplicate search. In this work a hand-drawn sketch highlighting the outline of an object of interest is used as query. Such a handdrawn
sketch places the focus on the shape of an object - rather than on appearance of a scene. For such
a task a common representation for the sketch and the images in the database is required. We propose
novel shape-based features calculated for local contour fragments. The shape features are stored in a
hierarchical data structure, which enables efficient retrieval in sub-linear time, scales to millions of images and does not require retraining when inserting new images. Our system is highly efficient and retrieves similar images in a database in a moderate 50 milliseconds on a standard single core computer.