Semantic segmentation of large-scale point cloud urban scenes faces significant challenges due to the complexity and diversity of object distribution. Transformer can effectively model the long-range dependencies in large-scale urban scenes, facilitating a comprehensive understanding of the overall structure and global context of urban scenes. Owing to the inherent strengths of the Transformer architecture, an increasing number of studies have applied it to the processing of large-scale point cloud data in complex scenes. But the excessive focus on global relationship inevitably leads to information redundancy, which can affect the model’s performance. To address this issue, we propose a novel Point Cloud Urban Scenes Semantic Segmentation network called PGFormer. This network consists of our proposed Group Representation Transformer (GRT) and KPConv, ensuring the enhancement of local key information while preserving global context. Specifically, in the proposed GRT block, we compute group-based and feature-based attention maps for Q and K across different groupings, and obtain the final attention output by integrating these with the global V, which has been pre-embedded with positional encoding derived from a triangular function. Our model is tested on two MLS (Paris-Lille-3D, Toronto3D) datasets and two ALS (Hessigheim 3D, ISPRS Vaihingen) datasets, and compared with a range of state-of-the-art (SOTA) methods, demonstrating the outstanding performance of our approach. Among various land cover segmentation tasks, our model achieves best or second-best results, particularly attaining the best OA (99.1%) and mF1-score (90.9%) on the Lille2 subset. Codes are available at https://github.com/Kange7/PGFormer.

Point cloud registration serves as a fundamental problem across multiple fields including computer vision, computer graphics, and remote sensing. While local feature descriptors (LFDs) have long been established as a cornerstone for point cloud registration and the LFD-based approach has been extensively studied, the field has witnessed significant advancements in recent years. Despite these developments, the research community lacks a systematic review to consolidate these contributions, leaving many researchers unaware of recent progress in LFD-based registration. To address this gap, we present a comprehensive review that critically examines both state-of-the-art and widely referenced methods across all subtasks of LFD-based registration. Our work provides: (1) an extensive survey of existing methodologies, (2) in-depth analysis of their respective strengths and limitations, (3) insightful observations and practical recommendations, and (4) a thorough summary of relevant applications and publicly available datasets. This systematic overview offers valuable guidance for researchers pursuing future investigations in this domain.

Point cloud semantic segmentation plays a key role in scene understanding and digital twin cities tasks. This article proposed a multi-granularity feature fusion network (MGF-Net) for point cloud semantic segmentation. The model first used a cluster relation aggregation module to extract fine-grained point features and a 3D convolution module to extract coarse-grained voxel features, followed by feature aggregation via a multi-granularity feature adaptive fusion module. Finally, to further improve the model performance, MGF-Net used a global feature attention module to capture long-distance context information. The performance of MGF-Net was evaluated on three point cloud datasets of urban scenes, i.e., Toronto3D, WHU-MLS, and SensatUrban. The quantitative results showed that MGF-Net achieved 80.16%, 51.27%, and 54.20% of mIoU on these datasets, respectively. Moreover, the comparative results showed that the proposed MGF-Net outperformed the baseline for complex urban scenes, and obtained better point cloud semantic segmentation results.

In recent years, fully supervised learning based semantic segmentation algorithms for point clouds have achieved significant advancements. However, a major limitation of these traditional algorithms is their reliance on extensive labeled datasets. This impedes their practical applicability. To overcome this obstacle, this paper proposes a novel point cloud semantic segmentation framework based on weakly supervised learning. This framework is designed to segment point cloud data both efficiently and accurately, even with a limited budget (0.1%) of labeled data. The proposed approach initiates with an active learning annotation strategy. This strategy involves computing the uncertainty scores of each point and ranking them, consequently selecting the top-K points for labeling based on the labeling budget. Furthermore, this paper developed a weakly supervised learning network. This network is enhanced by the calculation of multiple consistency losses to enhance the network's performance. Experimental results demonstrate that with a mere 0.1% labeling ratio, the proposed framework achieves a mean Intersection over Union (mIoU) of 70.3% on the NPM3D dataset.

Convolutional neural networks (CNNs) have achieved tremendous success in computer vision tasks, such as building extraction. However, due to domain shift, the performance of the CNNs drops sharply on unseen data from another domain, leading to poor generalization. As it is costly and time-consuming to acquire dense annotations for remote-sensing (RS) images, developing algorithms that can transfer knowledge from a labeled source domain to an unlabeled target domain is of great significance. To this end, we propose a novel full-level domain adaptation network (FDANet) for building extraction by combining image-, feature-, and output-level information effectively. At the input level, a simple Wallis filter method is employed to transfer source images into target-like ones whereby alleviating radiometric discrepancy and achieving image-level alignment. To further reduce domain shift, adversarial learning is used to enforce feature distribution consistency constraints between the source and target images. In this way, feature-level alignment can be embedded effectively. At the output level, a mean-teacher model is introduced to enforce transformation-consistent constraint for the target output so that the regularization effect is enhanced and the uncertain predictions can be suppressed as much as possible. To further improve the performance, a novel self-training strategy is also employed by using pseudo labels. The effectiveness of the proposed FDANet is verified on three diverse high-resolution aerial datasets with different resolutions and scenarios. Extensive experimental results and ablation studies demonstrated the superiority of the proposed method.

Traffic-sign recognition plays an important role in road transportation systems. This letter presents a novel two-stage method for detecting and recognizing traffic signs from mobile Light Detection and Ranging (LiDAR) point clouds and digital images. First, traffic signs are detected from mobile LiDAR point cloud data according to their geometrical and spectral properties, which have been fully studied in our previous work. Afterward, the traffic-sign patches are obtained by projecting the detected points onto the registered digital images. To improve the performance of traffic-sign recognition, we apply a convolutional capsule network to the traffic-sign patches to classify them into different types. We have evaluated the proposed framework on data sets acquired by a RIEGL VMX-450 system. Quantitative evaluations show that a recognition rate of 0.957 is achieved. Comparative studies with the convolutional neural network (CNN) and our previous supervised Gaussian-Bernoulli deep Boltzmann machine (GB-DBM) classifier also confirm that the proposed method performs effectively and robustly in recognizing traffic signs of various types and conditions.

Probabilistic registration algorithms [e.g., coherent point drift, (CPD)] provide effective solutions for point cloud alignment. However, using the original CPD algorithm for automatic registration of terrestrial laser scanner (TLS) point clouds is highly challenging because of density variations caused by scanning acquisition geometry. In this letter, we propose a new global registration method, introducing the use of the CPD framework for TLS point clouds. We first consider the measurement geometry and the intrinsic characteristics of the scene to simplify points. In addition to the Euclidean distance, we incorporate geometric information as well as structural constraints in the probabilistic model to optimize the so-called matching probability matrix. Among the structural constraints, we use a spectral graph to measure the structural similarity between matches at each iteration. The method is tested on three data sets collected by different TLS scanners. Experimental results demonstrate that the proposed method is robust to density variations and can decrease iterations effectively. The average registration errors of the three data sets are 0.05, 0.12, and 0.08 m, respectively. It is also shown that our registration framework is superior to the state-of-the-art methods in terms of both registration errors and efficiency. The experiments demonstrate the effectiveness and efficiency of the proposed probabilistic global registration.