Outdoor visual localization using aerial imagery as a map. From Master thesis to paper at CVPR 2023

In this upcoming CVPR 2023 paper, we address the task of cross-view pose estimation: given a ground-level camera image and an aerial patch of the local surroundings (think Google Maps), can we determine the exact pose (location + orientation) of the camera within the aerial view? This task is relevant for self-driving vehicles and outdoor robotics, especially when high-detailed maps are unavailable or outdated, and when GPS is inaccurate and only gives us a rough localization estimate. We present a new neural network design that incorporates geometric knowledge of the camera’s viewing frustum to efficiently compare visual features from the aerial image to the visual features in the ground-level camera image, setting a new state-of-the-art.

This work is a collaboration by Julian Kooij and Holger Caesar in the Intelligent Vehicles group at 3ME, both ELLIS Delft faculty. It is a great result of the MSc thesis of Ted de Vries Lentsch, who was supervised by Julian’s PhD student Zimin Xia. Congrats Ted and Zimin!

"SliceMatch: Geometry-guided Aggregation for Cross-View Pose Estimation", T. Lentsch, Z. Xia, H. Caesar, J.F.P. Kooij,
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, https://arxiv.org/abs/2211.14651