Accurate and Efficient Visual Relocalization
Title: Accurate and Efficient Visual Relocalization
SNIC Project: Berzelius-2022-15
Project Type: LiU Berzelius
Principal Investigator: Kunal Chelani <>
Affiliation: Chalmers tekniska högskola
Duration: 2022-02-01 – 2022-08-01
Classification: 10207


Visual relocalization is an important factor that can determine the performance of many mobile-robot-based applications such as virtual building tours, rescue operations, and even mixed reality experiences. The geometric approach to this problem is to first build a 3D representation of the environment using a set of database images within a Simultaneous Localization and Mapping (SLAM) or Structure from Motion (SfM) pipeline, and then use correspondences between 2D pixels in a query image and 3D points in the mapped environment to solve the Perspective-n-point problem, to produce a position and orientation from which the image was captured. Many recent localization pipelines establish matches between query and database images first and then use the 3D points corresponding to the matched 2D pixels in database images. State-of-the-art algorithms for image-matching use recent advancements in attention-based deep neural networks to produce robust matches between even in textureless regions of images. However, we believe that additional depth information form 3D points can be harnessed to produce rather accurate matches directly, circumventing the need to match images. This can lead to faster and more accurate localisation.