Dense 3D Reconstruction with a Hand-Held Camera

Abstract

Abstract. In this paper we present a method for dense 3D reconstruction from videos where object silhouettes are hard to retrieve. We introduce a close coupling between sparse bundle adjustment and dense multiview reconstruction, which includes surface constraints by the sparse point cloud and an implicit loop closing via the dense surface. The surface is computed in a volumetric framework and guarantees a dense surface without holes. We demonstrate the flexibility of the approach on indoor and outdoor scenes recorded with a commodity hand-held camera.

A Convex Relaxation Approach to Space Time Multi-view 3D Reconstruction

“…A similar model to the one in [8] has also recently been used by Ummenhofer and Brox [15] for combined 3D reconstruction and camera pose estimation. We adopt their approach because this model has several desirable properties.…”

A comparative study of video-based object recognition from an egocentric viewpoint

“…In the literature, [34] provides a comprehensive survey of how 3D CAD models can be used in contentbased retrieval systems. Several recent studies, including object recognition [35], landmark recognition [36] and camera pose estimation [37], have exploited 3D models reconstructed by photogrammetric methods, such as stereo matching [38] and structure-from-motion [39].…”

Fast dense 3D reconstruction using an adaptive multiscale discrete-continuous variational method

“…The camera pose refinement is achieved with a standard Bundle Adjustment involving all images. In [20], initial camera poses are refined iteratively with the feedback from optical flow between the texture on rendered surface and the actual images.…”

“…The results and computation statics of our system are uploaded in the evaluation page [1]. Because the smoothness constraint is used for the computation of depth maps, the reconstructed 3D models are smooth and complete without any optimization on voxel grid as in [23,22], or refinement of the obtained polygon model as in [10,20]. Our system is among the most efficient methods with 3 minutes run-time for each dataset.…”