We propose a comprehensive strategy to reconstruct urban building geometry from three-dimensional (3D) point clouds. First, the point clouds are segmented using a rough-detail segmentation algorithm, and refinements guided by topological relationships are performed to rectify the segmentation mistakes. Then, the semantic features (such as facades and windows) that belong to the buildings are recognized and extracted. Next, each facade is cut into a sequence of slices. The initial models are recovered by sequentially detecting and connecting the anchor points. Finally, due to the regular arrangements of windows, a template-matching method relying on the similarity and repetitiveness of the windows is proposed to recover the details on building facades. The experimental results demonstrate that our method can automatically reconstruct the building geometry and detailed window structures are better depicted.

Large-scale place recognition based on point clouds remains a challenge due to the geometric transformations in the scene. Among many geometric transformations, rotation is particularly challenging to handle in practice. Only a few works try to encode rotation invariance into learning based place recognition algorithms. In this letter, we propose a novel rotation-invariant aware network for large-scale place recognition named RIA-Net, which combines semantic and rotation-invariant geometric features to obtain discriminative and generalizable descriptors. First, we design seven dimensional rotation-invariant geometric features based on coordinate differences, distances, and angles in Euclidean space. Moreover, the semantic features in the local region are also considered to enrich the contextual information of scene point clouds. Furthermore, a feature augmentation module is developed to strengthen the rotation-invariant ability of the extracted features. Finally, a novel multi-scale spatial attention module is designed to establish long-range context dependencies for the local descriptors with different dimensions. Experiments on benchmark datasets demonstrate that the proposed RIA-Net outperforms the state-of-the-art place recognition methods. Compared to other rotation-invariant methods, RIA-Net can achieve better results in dealing with rotation problems.

Point cloud based place recognition remains a challenging issue due to the difficulty in extracting and encoding local features into robust and discriminative global descriptors from large-scale scenes. In this letter, we propose a novel deep neural network named SAGE-Net (Spatial Attention and Geometric Encoding Network) that incorporates both semantic and geometric features to generate discriminative and generalizable global descriptors. The spatial attention mechanism is employed to encode important inter-spatial relationships into local features. Then, the graph neural network is introduced to obtain the spatial distribution of features. Meanwhile, a geometric representation module is constructed by incorporating positional and angular information encoded in point coordinates. Then, a novel distance-based attentive pooling module is designed to extract informative local features based on the geometric distances between points. The description ability of global descriptors is improved by considering both contextual information among features and geometric features. Experiments on four benchmark datasets demonstrate that the proposed SAGE-Net outperforms the state-of-the-art approaches and shows good generalization ability for unseen scenes.