A Novel Depth Information-Guided Multi-View 3D Curve Reconstruction Method

Curve features, being more versatile than line segment features, are better suited for scene abstraction. However, obtaining three-dimensional (3D) curves with high scene coverage from multi-view images is a challenging task. In this study, we proposed a 3D curve reconstruction method guided by depth information. By utilizing depth information to narrow the search range of candidate two-dimensional (2D) curve matches, we mitigate the interference of non-corresponding curves on 2D curve matching, thereby improving the accuracy and recall rate of 2D curve matching. This results in reliable 3D curves with high scene coverage. For a curve on the reference image, we use depth information to project it onto the search image and design a purely geometric similarity measurement based on the projected curve to obtain 2D curve correspondences. Then, we generate 3D hypotheses based on the two-view matching results and design a robust geometric similarity measurement to obtain concise and reliable 3D curves from the redundant 3D hypotheses. Finally, we provide a curve-based bundle adjustment to achieve 3D curves with higher positional accuracy. We tested our method on five open-source datasets, demonstrating its effectiveness in generating 3D curves with high scene coverage, particularly in curved structure areas. Our method reconstructs 3.69 times more 3D lines on average than the best comparison method on five datasets, while also achieving higher positional accuracy.