Understanding the Limitations of SfM-Based Camera Calibration on Multi-View Stereo Reconstruction

The recent success on convolutional neural networks (CNNs) research has triggered the interest to improve the stereo reconstruction. There have been many considerations to reconstruct unstructured general images. However, photos taken in an unstructured environment are not well applied to previous research techniques. This paper proposes a robust method in this special environment. First, camera parameters are extracted using two types of Structure-from-Motion (SfM). In general, intrinsic camera parameters are extracted via camera calibration, and extrinsic parameters are computed by SfM. After that, we put these parameters as an input into two types of multiview stereo methods and compare the results of each methods. Finally, the structured indoor dataset which is called ETRI 360° dataset has better reconstruction results compared with unstructured dataset by using specific SfM and MVS methods.