3D Reconstruction from an Uncalibrated Long Image Sequence

Abstract

In this paper, we propose a seamless and dense 3D reconstruction method that uses an uncalibrated long image sequence. First, we divide a long image sequence into subsequences. Camera motion parameters are estimated by a factorization method optimized for near planar motion from each subsequence, because we can assume that camera motion will be near planar motion in short sequences. Dense 3D points are reconstructed by applying a multi view stereo technique to each subsequence using the estimated camera motion parameters. We then find the relationships between the subsequences from the overlapping frames of adjacent subsequences. Finally, seamless and dense 3D reconstruction is achieved by merging all 3D points obtained from each subsequence. Experiments conducted on a long image sequence captured by a handheld camera while walking demonstrate that the proposed method yields good results.