The complexity and variety of buildings and the defects of point cloud data are the main challenges faced by 3D urban reconstruction from point clouds, especially in metropolitan areas. In this paper, we developed a method that embeds multiple relations into a procedural modelling process for the automatic 3D reconstruction of buildings from photogrammetric point clouds. First, a hybrid tree of constructive solid geometry and boundary representation (CSG-BRep) was built to decompose the building bounding space into multiple polyhedral cells based on geometric-relation constraints. The cells that approximate the shapes of buildings were then selected based on topological-relation constraints and geometric building models were generated using a reconstructing CSG-BRep tree. Finally, different parts of buildings were retrieved from the CSG-BRep trees, and specific surface types were recognized to convert the building models into the City Geography Markup Language (CityGML) format. The point clouds of 105 buildings in a metropolitan area in Hong Kong were used to evaluate the performance of the proposed method. Compared with two existing methods, the proposed method performed the best in terms of robustness, regularity, and topological correctness. The CityGML building models enriched with semantic information were also compared with the manually digitized ground truth, and the high level of consistency between the results suggested that the produced models will be useful in smart city applications.

Aerial laser scanning or photogrammetric point clouds are often noisy at building boundaries. In order to produce regularized polygons from such noisy point clouds, this study proposes a hierarchical regularization method for the boundary points. Beginning with detected planar structures from raw point clouds, two stages of regularization are employed. In the first stage, the boundary points of an individual plane are consolidated locally by shifting them along their refined normal vector to resist noise, and then grouped into piecewise smooth segments. In the second stage, global regularities among different segments from different planes are softly enforced through a labeling process, in which the same label represents parallel or orthogonal segments. This is formulated as a Markov random field and solved efficiently via graph cut. The performance of the proposed method is evaluated for extracting 2D footprints and 3D polygons of buildings in metropolitan area. The results reveal that the proposed method is superior to the state-of-art methods both qualitatively and quantitatively in compactness. The simplified polygons could fit the original boundary points with an average residuals of 0.2 m, and in the meantime reduce up to 90% complexities of the edges. The satisfactory performances of the proposed method show a promising potential for 3D reconstruction of polygonal models from noisy point clouds.

Extracting rocks from Mars rover images using convolutional neural networks is a semantic segmentation problem gaining increasing attention in planetary science and artificial intelligence. However, this task still faces the challenges of inaccurate extraction of rock boundaries and small rocks. An important reason is that the textural features of the Mars rover images are not discriminative enough between the target and the background. In order to obtain 3-D information to enhance the rock extraction without introducing additional data, we designed a dual-task branch network with a Y-shaped encoder-decoder structure. In our network, the primary semantic segmentation task branch is used to decode textural features into semantic features and output rock extraction results; the auxiliary task depth estimation branch decodes textural features into depth features in 3-D, and transmits them to the primary branch through the spatial attention module to enhance the identification ability of semantic features for boundaries and small targets. In addition, another reason impedes accurate Mars rock extraction is the lack of high-quality annotated training datasets. Therefore, we created a dataset containing 6,325 images pairs with corresponding annotations and depth information, SimMars6K. Ablation and comparison experiments based on this dataset and two actual datasets show that our method achieves 83.1% IoU and 90.8 F ₁ -score on simulated dataset, and outperforms other methods on actual datasets with 2% improvement in recall. Transfer learning experiment shows that the pre-training on our simulated data can bring up to additional 5% gain on mIoU and 6% gain on mPA for general model.