📝 SummarXiv

Abstract

We propose an accurate and highly interpretable fine-grained cross-view localization method that estimates the 3 Degrees of Freedom pose of a ground-level image by matching its local features with a reference aerial image. Previous methods typically transform the ground image into a bird's-eye view (BEV) representation and then align it with the aerial image for localization. However, this transformation often leads to information loss due to perspective distortion or compression of height information, thereby degrading alignment quality with the aerial view. In contrast, our method directly establishes correspondences between ground and aerial images and lifts only the matched keypoints to BEV space using monocular depth prior. Notably, modern depth predictors can provide reliable metric depth when the test samples are similar to the training data. When the depth distribution differs, they still produce consistent relative depth, i.e., depth accurate up to an unknown scale. Our method supports both metric and relative depth. It employs a scale-aware Procrustes alignment to estimate the camera pose from the correspondences and optionally recover the scale when using relative depth. Experimental results demonstrate that, with only weak supervision on camera pose, our method learns accurate local feature correspondences and achieves superior localization performance under challenging conditions, such as cross-area generalization and unknown orientation. Moreover, our method is compatible with various relative depth models without requiring per-model finetuning. This flexibility, combined with strong localization performance, makes it well-suited for real-world deployment.