China
Africa
Explore chapters and articles related to this topic
Multiview Image Matching for 3D Earth Surface Reconstruction
Published in Yuhong He, Qihao Weng, High Spatial Resolution Remote Sensing, 2018
Regarding the definition of fs, photo consistency is the essential and most popular similarity measure in optical image matching. There are many similarity measures listed in Table 5.1, such as pixel-based window matching and sum of squared differences, normalized cross-correlation (NCC) (Hannah, 1974), gradient-based measures (Scharstein, 1994), image variance and entropy (Zitnick et al., 2004), and hierarchical mutual information (Hirschmuller, 2008). According to Hirschmuller & Scharstein (2009), similarity measures (or cost or distance measures in the opposite manner) roughly fall into three categories: (a) parametric measures based on the magnitude of image intensities, (b) nonparametric measures based on the local order of image intensities, and (c) mutual information measures based on relationships between images. Detailed summaries of similarity measures are given in Brown et al. (2003) and Hirschmuller & Scharstein (2009).
The role of machine intelligence in photogrammetric 3D modeling – an overview and perspectives
Published in International Journal of Digital Earth, 2021
The development of image-based per-pixel dense matching for DSM generation has been quite significant in the past decades and many techniques were developed both in the photogrammetry (Gruen, 2012) and computer vision community, e.g. multi-image and constraint-based matching (Zhang, 2005), dynamic programming (Veksler, 2005), semi-global matching (Hirschmüller, 2008), patch-based matching (Furukawa and Ponce, 2010), graph-cut (Vicente et al., 2008). Dense matching can be generally categorized based on the number of images used for computation: (1) stereo matching and (2) multi-view image matching. This leads to some fundamental algorithmic differences (Broadhurst et al., 2001; Furukawa et al., 2010; Furukawa and Ponce, 2009; Zhang, 2005): for example, stereo matching is able to utilize the rectified epipolar images for easier algorithm implementation (correspondences lying on the same row), while epipolar rectified images do not generally exist for more than two images. An obvious advantage of multi-view matching is that it is able to take redundant measurements to improve the robustness and accuracy of per-point matching (Zhang and Gruen, 2006). A few algorithms have shown great advances in utilizing the multiple observations such as multi-photo matching (Baltsavias, 1991), patch-based multi-view matching (Furukawa and Ponce, 2010), voxel-based space carving (Broadhurst et al., 2001). The state-of-the-art algorithms formulate the photo-consistency condition (computed from two or more images) within a global energy optimization framework. Points matched through multi-image matching provide the optimal accuracy, while in practice if the occluded pixels (both in multi-image and stereo matching) are not handled carefully, it may pose negative impact on neighboring pixels through the solver. For instance, the object space semi-global matching (SGM) algorithm (Bethmann and Luhmann, 2015) takes the average matching scores across multiple images, and turns the disparity computation procedure in a voxel object space. Since a mechanism determining the occluded pixels before averaging the scores is lacking, the results are not reported better than in the original algorithms. Given that stereo-based matching algorithms are particularly effective, practical implementation sometimes favors a multi-depth fusion approach (multi-stereo algorithms) (Seitz et al., 2006; Wenzel et al., 2013), where stereo matching are performed on permutated and selective pairs and then a depth/DSM fusion step utilizes the redundant information in the object space. This type of methods leaves the information fusion in the object space and can easily extend the state-of-the-art stereo matching algorithms to multi-view scenario. However, a theoretically more powerful concept is that of geometrically constrained multi-view matching. It allows to determine the matching parameters for all images involved plus the object space coordinates of the point in question in one simultaneous solution. By computation of the covariance matrix of all system unknowns one has an excellent tool for quality analysis of the matching process. Depending on the situation different constraints can be formulated: Epipolar-; collinearity-; X,Y-; Z-constraint.