Abstract

Location recognition is the problem of determining the place depicted in a given photo. In the first part of the talk, we consider the case where the scene is represented by a 3D model and we are not only interested in the place depicted in an image but also the position and orientation from which the image was taken, i.e., the camera pose. A major challenge for solving this image-based localization problem is to establish the 2D-3D matches required for pose estimation. This is especially true for large scale scenes containing many 3D points with similar local appearance. Instead of using elaborate matching schemes, we introduce an efficient camera voting approach whose run-time is independent of the inlier ratio. While our approach allows us to handle an arbitrary number of matches in linear time, we show that simply increasing the number of 2D-3D matches used for pose estimation does not solve the image-based localization problem. The reason for this behavior is that increasing the

number of matches leads to wrong poses with many inliers. In the second part of the talk, we thus consider the problem of finding a better decision criterion than the raw inlier count. We show that geometric bursts, i.e., spatial configurations appearing at multiple places in a scene, are a major reason why the raw inlier count fails for large-scale location recognition. We introduce simple schemes that allow us to efficiently detect geometric bursts during query time. We show experimentally that down-weighting inliers based on the number of bursts they appear in allows us to better decide between correct and incorrect place recognition results and significantly boosts the location recognition performance.