Sketch Based Retrieval with Asymmetric Feature Maps

Giorgos Tolias
(Center for Machine Perception, CTU Prague, Czech Republic)


In this work we propose a short vector image representation that supports efficient scale and translation invariant sketch-based image retrieval. The efficiency of the search is boosted by two means: approximating a trigonometric polynomial of scores and by decomposing a 2D translation search by 1D projections. We introduce a novel concept of asymmetric feature maps, which allows to evaluate multiple kernels without increasing memory requirements. This enables multi-scale search as well as translation approximation by projections. The representation is learned by joint feature maps optimization.