Open source stack for Structure from Motion 3D reconstruction: A geometric overview

摘要

Given a series of overlapped images, it is possible to reconstruct camera position and orientation, as well as 3D coordinates of object in the images. Such method is known as Structure from Motion (SfM), which is currently implemented in various commercial and open source software. The first software commonly don't provide Sfm algorithm. While the latter software are commonly developed partially. This research aims to perform a SfM reconstruction using a stack of open source SfM software, with each software serves different purposes in producing 3D models from images. A python script were employed to produce seamless integration of the software. Geometric evaluations were performed to the result of SfM reconstruction on three datasets obtained using mobile handheld camera. The deliverables of the stack were compared to result of commercial software using cloud-to-cloud comparison to obtain geometric closeness of both the results. Ground truth measurement were conducted to assess the positional consistency in SfM reconstruction on different dataset. The test showed that 3D models resulted from the open source stack perform on pair with the result of Agisoft Photoscan as indicated by the small RMSE value obtained from cloud-to-cloud comparison (0.019818 m and 0.350701 m for two datasets, respectively). The result of ground-truth evaluation indicates large error possibly due to the homogeneity in the dataset used, as indicated by the camera pose graph.