3D curve sketch: Flexible curve-based stereo reconstruction and calibration

Abstract

Interest point-based multiview 3D reconstruction and calibration methods have been very successful in select applications but are not applicable when an abundance of feature points are not available. They also lead to an unorganized point cloud reconstruction where the geometry of the scene is not explicit. The multiview stereo methods on the other hand yield dense surface geometry but require a highly controlled or calibrated setting. We propose and develop a novel framework for 3D reconstruction and calibration based on image curve content, whose output is a 3D curve sketch, an unorganized set of 3D curve fragments. This approach, which is meant to augment the previous approaches, results in a reconstruction of geometric curve structure which can serve as a scaffold on which surface patches can be potentially reconstructed. It is intented for the setting where a number of images are available with coarsely calibrated cameras. The approach operates in two stages. A reliable partial 3D curve sketch is first reconstructed and this is used to refine the cameras to yield a more complete 3D curve sketch in a second stage. A key advantage of this approach is the ability to integrate information across a large number of views. The results have been evaluated on a few datasets.

Image-based reconstruction of wire art

“…This process typically results in multiple ambiguous matches and an additional view is required for verification (Robert and Faugeras 1991). Recent work (Fabbri and Kimia 2010;Usumezbas et al 2016) uses this workflow to reconstruct curvilinear structures in general scenes requiring a dense set of images. Our work, in contrast, reconstructs wireframe objects from as few as three input views.…”

2-manifold reconstruction from sparse visual features

“…Unfortunately, curve geometry based matching method as described in [3] exhibited bad performance when the images contained a lot of condensed and similar curves (for example, as in the Ladybug datasets, see Section 3), it just can't make a matching decision based on curve geometry alone. So, we decided to use a correlation based method heavily inspired by a classical points matching.…”

“…The work [4,11,16,17] (including ours [17]) on 2-manifold reconstruction from sparse SfM data does not do that. Our curve reconstruction method involves matching based on correlation and assume that the epipolar constraint is known (this is different from curve reconstruction in [3] where the correlation is not used). Section 2 summarizes the surface reconstruction method and describes the curve reconstruction.…”

iModel: Interactive Co-segmentation for Object of Interest 3D Modeling

“…However, as we show through our results (Section 4), our approach might be a good first-approximation for consumer applications. Recently, Fabbri et al [38] developed an algorithm which allows for reconstructing curves in the structure. Although not done here, this approach could provide cues to help tackle the problem of concavities in the structure.…”