133 / 2024-08-22 20:24:56

Evaluation of the Effect of Single Image Super Resolution on Stereo Matching from High Resolution Satellite Imagery AITC 2024+空天之星

Super Resolution, Stereo Matching, Remote Sensing Imagery, 3D Recon-struction, Deep Learning

3D reconstruction using very-high resolution (VHR) satellite remote sensing images has become one of the key researches in photogrammetry in recent years. As the demand for the construction of finer 3D models intensifies, the need for higher resolution satellite images has increased. However, the cost of enhancing satellite resolution through hardware improvement is prohibitively high. Consequently, single-image super-resolution (SISR) technology, which recovers the detailed information of high resolution (HR) images from given low resolution (LR) images, has become a mainstream approach. Despite this, there have been few studies on the combined application of SISR technology and 3D reconstruction, leading to a lack of experimental validation especially on VHR satellite images. In this paper, we conducted an extensive experimental evaluation of SISR and stereo matching using several large public remote sensing datasets, employing more than ten mainstream deep learning-based SISR methods alongside the prominent semi-global matching (SGM) algorithm. We systematically assess the impact of SISR on stereo matching. The experimental results indicate that the SISR techniques can produce 3D reconstruction results from SR images that are comparable to those obtained from original HR images to a certain extent. Additionally, no significant correlation is found between mainstream SISR evaluation metrics and the stereo matching accuracies.