FEATS: Synthetic Feature Tracks for Structure from Motion Evaluation

Joseph DeGol, Jae Yong Lee, Rajbir Kataria, Daniel Yuan, Timothy Bretl, Derek Hoiem

Abstract: We present FEATS (Feature Extraction and TrackingSimulator), that synthesizes feature tracks using a cam-era trajectory and scene geometry (e.g. CAD, multi-viewstereo). We introduce 2D feature and matching noise mod-els that can be controlled using a few parameters. We alsoprovide a new dataset of images and ground truth camerapose. We process this data (and a synthetic version) withseveral current SfM algorithms and show that the synthetictracks are representative of the real tracks. We then showtwo practical uses of FEATS: (1) we generate hundreds oftrajectories with varying noise and show that COLMAP ismore robust to noise than OpenSfM and VisualSfM; and (2)we calculate 3D point error and show that accurate camerapose estimates do not guarantee accurate 3D maps