Explore chapters and articles related to this topic
Visual and virtual progress monitoring in Construction 4.0
Published in Anil Sawhney, Mike Riley, Javier Irizarry, Construction 4.0, 2020
Jacob J. Lin, Mani Golparvar-Fard
The PMVS (Furukawa et al., 2009) algorithm consists of three steps: match, expand, and filter. Utilizing the camera information from the SfM process, a set of new features are detected and matched among all the images yielding 3D information of a sparse set of patches with important regions of the images. It expands the initial matches to the nearby pixels to generate denser patches around the initial patch and then filter out the incorrect matches to reduce the noises. Although PMVS is widely used, MVE (Fuhrmann et al., 2014) aims to generate a denser point cloud for multi-scale geometry mesh reconstruction to preserve more details in the result. It uses the depth map from every view to reconstruct the dense point cloud, even though the redundancy of many views are overlapped with similar parts, this helps to only reconstruct the other small parts of the scene that are not visible. Also, the depth map can be directly embedded with all the parameterized information such as color. After the MVS from MVE, to merge the depth map into a globally consistent geometry, it uses hierarchical Signed Distance Fields (SDF) to interpolate and approximate points based on sample scale and redundancy. The final result produces a globally consistent surface mesh. COLMAP (Schönberger and Frahm, 2016, Schönberger et al., 2016) further improved the MVS process through embedding pixelwise depth and normal estimation with a photometric and geometric prior restricted view selection. The PatchMatch stereo is performed in four different directions and optimized through generating normal and depth candidates from the previous iteration to perform the patch based matching. With a multi-view geometric consistency term developed, the integration can simultaneously refine and fuse the image depth and normal. After the final step to filter remaining outliers that are not photogrammetrically and geometrically stable with support from multiple views, the dense point cloud is colored and generate with normal. The MVS process results in a dense point cloud that is necessary for construction purposes such as occupancy analysis, material recognition, and safety inspection.
Image completion via texture direction analysis
Published in Journal of Modern Optics, 2019
Qiaochuan Chen, Guangyao Li, Li Xie, Qingguo Xiao, Mang Xiao
The PatchMatch (16) method is adopted to accelerate our algorithm when searching for the best matching patches. To reduce colour deviation, in this study, following HaCohen et al. (26), bias and gain are added. We restrict the bias/gain compensation with predefined ranges: [−50, 50] for bias, and [0.5, 1.5] for gain. Because the objective function consists of three items, our method requires more iterations to converge to an optimal value. Updating missing pixel
Structure constrained image completion by geometric transformation model and dynamic patches
Published in Journal of Modern Optics, 2019
Qiaochuan Chen, Guangyao Li, Li Xie, Qingguo Xiao, Mang Xiao
In the search step, the PatchMatch method (17) is used, which effectively finds the best matching patch for the target patch. This method is extended by using affine transformations rather than translations, rotations and scaling transformations. Extending the patch search space is helpful to obtain the optimal solution. Augmented space is significantly larger than that for just translation. Therefore, more similar patches can be found in larger space. We arrange more iterations for convergence to acceptably good results.