Establishing consistent and dense correspondences across multiple images is crucial for Structure from Motion (SfM) systems. Significant view changes, such as air-toground with very sparse view overlap, pose an even greater challenge to the correspondence solvers. We present a novel optimization-based approach that significantly enhances existing feature matching methods by introducing geometry cues in addition to color cues. This helps fill gaps when there is less overlap in large-scale scenarios. Our method formulates geometric verification as an optimization problem, guiding feature matching within detector-free methods and using sparse correspondences from detectorbased methods as anchor points. By enforcing geometric constraints via the Sampson Distance, our approach ensures that the denser correspondences from detector-free methods are geometrically consistent and more accurate. This hybrid strategy significantly improves correspondence density and accuracy, mitigates multi-view inconsistencies, and leads to notable advancements in camera pose accuracy and point cloud density. It outperforms state-ofthe- art feature matching methods on benchmark datasets and enables feature matching in challenging extreme largescale settings. Project page: https://xtcpete.github.io/geo-website

Correspondence, as perhaps the most fundamental problem in computer vision, helps us with mapping the world. Solving correspondence is required in most applications in Augmented Reality and Robotics since it derives the camera poses and 3D point locations [16]. In the 3D reconstruction from a set of uncalibrated images, the correspondence problem meets two additional demands: 1) the demand to be able to observe a corresponding point across multiple views, which is known as the track length of a correspondence point, and 2) the demand to find a dense set of such points such that they are sufficient for the recovery of 3D geometry of the scene. In this paper, we consider a very difficult correspondence problem—dense, consistent correspondences across a sequence of frames with large camera baselines. To our knowledge, no prior works have achieved satisfactory results in the presence of these challenges: both detectorbased and detector-free methods struggle with significant view changes, as illustrated in Figure 2, due to their inherent limitations. Detector-based methods often match only a small number of keypoints, making accurate image registration difficult.

We propose the geometry-aware optimization module, an innovative optimization approach that bridges detectorbased matcher and detector-free matcher to improve SfM reconstruction and accuracy of estimated camera poses. This module leverages epipolar constraints, iteratively optimizes for the best relative pose, and applies Sampson distance weighting to refine match precision. The introduction of geometry optimization enhances the system’s ability to handle challenging scenarios, including drastic changes in perspective, scale, and illumination. Our method is shown to be also effective in 3D reconstruction for large baselines, a scenario where conventional methods often fail and perform unsatisfied.

We propose the geometry-aware optimization module, an innovative optimization approach that bridges detectorbased matcher and detector-free matcher to improve SfM reconstruction and accuracy of estimated camera poses. This module leverages epipolar constraints, iteratively optimizes for the best relative pose, and applies Sampson distance weighting to refine match precision. The introduction of geometry optimization enhances the system’s ability to handle challenging scenarios, including drastic changes in perspective, scale, and illumination. Our method is shown to be also effective in 3D reconstruction for large baselines, a scenario where conventional methods often fail and perform unsatisfied.