Leveraging Deep Visual Descriptors for Hierarchical Efficient\n Localization

Abstract

Many robotics applications require precise pose estimates despite operating\nin large and changing environments. This can be addressed by visual\nlocalization, using a pre-computed 3D model of the surroundings. The pose\nestimation then amounts to finding correspondences between 2D keypoints in a\nquery image and 3D points in the model using local descriptors. However,\ncomputational power is often limited on robotic platforms, making this task\nchallenging in large-scale environments. Binary feature descriptors\nsignificantly speed up this 2D-3D matching, and have become popular in the\nrobotics community, but also strongly impair the robustness to perceptual\naliasing and changes in viewpoint, illumination and scene structure. In this\nwork, we propose to leverage recent advances in deep learning to perform an\nefficient hierarchical localization. We first localize at the map level using\nlearned image-wide global descriptors, and subsequently estimate a precise pose\nfrom 2D-3D matches computed in the candidate places only. This restricts the\nlocal search and thus allows to efficiently exploit powerful non-binary\ndescriptors usually dismissed on resource-constrained devices. Our approach\nresults in state-of-the-art localization performance while running in real-time\non a popular mobile platform, enabling new prospects for robotics research.\n