Guided Quasi-Dense Tracking for 3D Reconstruction

In this paper we propose a framework for obtaining quasi-dense Euclidean structure reconstruction by means of guided quasi-dense point tracking in image sequences. We use a stratified algorithm that establishes most reliable sparse point correspondences first then robustly estimates multiview geometry and finally propagates sparse point features to quasi-dense matches and tracks. Notable properties of our algorithm are built-in motion model selection using geometric robust information criteria (GRIC) that helps to avoid common algorithm degeneracy. Matching outliers are segmented by robust homography estimation for rotational camera movement or by reprojection error thresholding otherwise. We employ automatic keyframe selection that is used to provide frames with reasonable disparity for reliable matching and reconstruction. These technique also reduces cost of computationally expensive quasi-dense tracking.