Title:Multistage SFM: A Coarse-to-Fine Approach for 3D Reconstruction

Abstract:Several methods have been proposed for large-scale 3D reconstruction from large, unorganized image collections. A large reconstruction problem is typically divided into multiple components which are reconstructed independently using structure from motion (SFM) and later merged together. Incremental SFM methods are most popular for the basic structure recovery of a single component. They are robust and effective but are strictly sequential in nature. We present a multistage approach for SFM reconstruction of a single component that breaks the sequential nature of the incremental SFM methods. Our approach begins with quickly building a coarse 3D model using only a fraction of features from given images. The coarse model is then enriched by localizing remaining images and matching and triangulating remaining features in subsequent stages. These stages are made efficient and highly parallel by leveraging the geometry of the coarse model. Our method produces similar quality models as compared to incremental SFM methods while being notably fast and parallel.