Improved SfM-MVS approach using artificial backgrounds

Abstract

With the development of computer vision technology, structure from motion and multiview-stereo (SfM-MVS) approach has been widely applied in the geotechnical field. However, as a method that utilizes a series of images to reconstruct a 3D model, errors often occur due to insufficient feature points in the images. In this study, soil blocks, rubber specimens, and a sand particle ranging in size from 10 cm to 0.3 mm were utilized for synthetizing 3D model by the SfM-MVS approach. Additionally, an artificial background containing various colored blocks was introduced during photographing process to improve this approach. Moreover, the application of this approach was extended to process optical microscope images and scanning electron microscope (SEM) images by using two artificial backgrounds. Experimental comparison suggested that using artificial backgrounds could optimize the depression areas between the specimen and sample holder of the three-dimensional (3D) model generated by the SfM-MVS approach, especially in the depression portions with acute angles. And the reconstructed model from the SfM-MVS approach was comparable to that generated by X-ray computed tomography (CT). It was also found that increasing the image resolution and decreasing voxel size can improve the accuracy of the 3D model. And these improvements have been quantitatively demonstrated by tests. When using optical microscopy and SEM, the application of artificial backgrounds significantly increased the success rate of constructing 3D models, compared to the near impossibility of achieving successful reconstruction without them in the practice. It was mainly attribute to sufficient feature points in artificial backgrounds can be captured from artificial backgrounds in the camera tracking and point-matching processes of the SfM-MVS approach. With the proposed method in this study, the applicability of the SfM-MVS approach was extended in laboratory geotechnical experiments.

Keywords

• Structure from motion
• Multiview-stereo approach
• Particle shape
• 3D model
• Scanning electron microscope