Relative Feature Orientation Filtering in COLMAP Structure from Motion

Steven Mills

Abstract: Structure-from-motion and multi-view stereo systems are based on feature correspondences between images. While these features are typically treated as points, common feature detection/description methods such as SIFT provide orientation and scale as well as location information. Previous work has shown that orientation information in particular can be exploited to improve feature matching and relative pose computation. This work examines the effect of using feature orientation information in the open-source COLMAP structure-from-motion software.. Experiments show that using feature orientation can reduce the time needed for both feature matching and initial scene reconstruction. Furthermore the resulting reconstructions have a lower average reprojection error with orientation-based filtering.

External IDs:dblp:conf/ivcnz/Mills23