Structure-from-Motion from Pixel-wise Correspondences

We introduce DMap, a Structure-from-Motion system that efficiently leverages dense, pixel-wise correspondences across images.Existing approaches like COLMAP typically rely on sparse correspondences and are limited by the lack of accuracy and repeatability of keypoint detection, especially in texture-less environments.Dense correspondences overcome these limitations but their sheer numbers cannot be effectively handled by these systems.Differently, DMap exploits the high parallelism of GPUs to handle hundreds of millions of correspondences in each image pair.As a result, DMap is significantly more robust and accurate than existing approaches, yet can scale to large scenes with thousands of images.Our experiments show that DMap results in better camera poses and denser reconstructions from fewer input views.DMap will be released with a permissive license.