A Benchmark for Visual Positioning from Depth Images

In the last years, there has been an increasing interest in research on visual positioning and localization. Several datasets have been published recently. However, little attention has been paid to architectural aspects of scenes that have the advantage of being not influenced by scenery changes. Based on depth videos recorded using depth sensor, we propose in this paper a new dataset composed by descriptors of 2D maps associated to computed 3D structures. We show also how to interrogate the dataset using a set of depth frames acquired in visited places. Conducted experiments show the accuracy of the obtained 2D maps and feasibility of the proposed framework.