• Two publicly available realistic 3D semantic change detection datasets. • A novel multi-task learning architecture for large-scale 3D point cloud semantic change detection. • A multi-dimensional change encoding module that refines the cross-temporal neighborhood variations. • A change-guided semantic refinement module that enhances the representation of semantic features. • A semantic-awareness change interaction module that complements the characteristic of change. 3D semantic change detection enables the detection and identification of both changes in urban objects and their semantic categories, providing fine-grained change information for downstream applications. Existing methods rely on single-branch architectures with predefined output labels, which is simple but suffers from complex output definitions and ineffective multi-task coupling, compounded by scarce annotated 3D realistic data. To overcome these challenges, firstly, two 3D realistic semantic change detection datasets are constructed and published, named HKSCD and UtrechtCD , which utilize oblique photogrammetry point clouds and LiDAR point clouds to describe 9 semantic categories and 2 change types in Hong Kong, China, and Utrecht, Netherlands, covering 15 square kilometers with 370 million points. Secondly, a Multi-task Interaction Siamese Network (MISNet) for 3D point cloud semantic change detection is proposed. It deeply couples semantic segmentation and change detection, enabling the simultaneous prediction of both tasks within a unified architecture. The proposed multi-dimensional change encoding module computes cross-temporal neighborhood relationships from multiple dimensions to extract accurate point cloud change features. Additionally, the change-guided semantic refinement module and the semantic-awareness change interaction module leverage change information to support semantic consistency and utilize semantic information to assist inter-class change detection to promote cross-task consistent modeling underpinned by the cross-learning strategy. Extensive experiments demonstrate that MISNet achieves mIoU of 84.15% (HKSCD), 85.15% (UtrechtCD), and 89.58% (Urb3DCD-V2), outperforming existing methods by + 2.21%, +1.43%, and + 1.46%, respectively. The code and dataset are available at https://github.com/zhanwenxiao/UrbanSCD and https://github.com/zhanwenxiao/MISNet .

Point cloud instance segmentation for building indoor scenes using deep learning and BIM-Generated synthetic point clouds

We propose Next Bit Prediction (NBP), a unified framework that simultaneously addresses lossless compression and lossy reconstruction of 3D point cloud geometry through a next-bit probability estimation paradigm. Our key insight is that both lossless compression and lossy reconstruction fundamentally rely on accurate probability estimation of geometric symbols, though targeting different metrics. Lossless compression minimizes bitrate via precise symbol distribution prediction, while lossy reconstruction enhances reconstruction fidelity through probability-guided geometry refinement. Recognizing that point clouds become sparser with increasing bit depth, NBP introduces two key technical innovations. For more significant bits, where the point density is higher, we develop a multi-stage Occupancy Probability Estimation (OPE) mechanism to estimate the probability distribution of occupancy status across multiple iteration stages, with each stage supporting either lossless or lossy mode. For less significant bits that focus on point placement, a Disentangled Probability Estimation (DPE) module is proposed to handle density information and binary residuals, simultaneously enabling lossless compression and facilitating probability-driven coordinate refinement for high-quality lossy reconstruction. Extensive experiments demonstrate the advantages of NBP, including low complexity, progressive coding, and superior coding efficiency, achieving state-of-the-art results both quantitatively and qualitatively.

AI-driven crown generation: A comparative analysis of point cloud completion models for mandibular first molar restoration.

Safety-critical control using high-dimensional sensory feedback from optical data (e.g., images, point clouds) poses significant challenges in domains like autonomous driving and robotic surgery. Control can rely on low-dimensional states estimated from high-dimensional data. However, the estimation errors often follow complex, unknown distributions that standard probabilistic models fail to capture, making formal safety guarantees challenging. In this work, we introduce a novel characterization of these general estimation errors using sub- Gaussian noise with bounded mean. We develop a new technique for uncertainty propagation of proposed noise characterization in linear systems, which combines robust set-based methods with the propagation of sub-Gaussian variance proxies. We further develop a Model Predictive Control (MPC) framework that provides closed-loop safety guarantees for linear systems under the proposed noise assumption. We apply this MPC approach in an ultrasound-image-guided robotic spinal surgery pipeline, which contains deep-learning-based semantic segmentation, image-based registration, high-level optimization-based planning, and low-level robotic control. To validate the pipeline, we developed a realistic simulation environment integrating real human anatomy, robot dynamics, efficient ultrasound simulation, as well as in-vivo data of breathing motion and drilling force. Evaluation results in simulation demonstrate the potential of our approach for solving complex image-guided robotic surgery task while ensuring safety.

Close-range airborne LiDAR captures high-density, high-accuracy point clouds, offering strong potential for spatially explicit Leaf Area Index (LAI) estimation in complex forests. However, key challenges in the modelling process — particularly LiDAR detection rates and accounting for the wood component — remain inadequately explored. Using simulated UAV LiDAR data generated with the Discrete Anisotropic Radiative Transfer (DART) model and realistic forest mock-ups, we evaluate how these factors influence LAI estimates. Simulated point clouds are processed with the open-source ray-tracing tool AMAPVox to assess (1) the bias introduced by incomplete detection and (2) the effectiveness of using a wood mask to exclude woody contributions. Our results show that pulse fragmentation promotes incomplete detection, which biases LAI estimates; the magnitude of this bias depends on the return-weighting strategy. For example, removing returns with small reflectance, which collectively contribute less than 5% of the backscattered energy, results in an overestimation of LAI by 10% to 20%. An alternative very simple and highly scalable weighting strategy consisting of selecting the strongest return per pulse is also explored and is shown to be effective in most cases albeit slightly less accurate. Additionally, segmentation accuracy declines as the proportion of mixed leaf-wood points increases. These findings suggest that small footprint LiDAR systems are better suited for LAI mapping: their smaller footprints reduce beam fragmentation (mitigating echo-weighting uncertainties) and minimize mixed points while improving geolocation accuracy — critical for robust leaf/wood classification. These insights advance best practices for LiDAR-based forest monitoring.

LU-mamba: LiDAR upsampling via bidirectional state space modeling on range images.

Monitoring of segmental lining uplift and synchronous grouting control during shield tunnel construction using point-cloud data

Biofouling-covered bridge pile morphology assessment based on underwater sonar point cloud data

The stability assessment and potential sliding surface determination of open-pit mine slopes are pivotal topics in mine safety. Focusing on an open-pit limestone mine slope in Hebei Province, China, this study proposes a systematic and practical methodology for modeling, stability analysis, and critical sliding surface determination of three-dimensional (3D) slopes in engineering projects. A 3D grid model of mine slope characterized by a complex surface configuration was constructed through a series of meticulous operations using high-quality point cloud data. The safety factor and critical failure state of the slope were determined using a compressive strength reduction strategy based on the generalized Hoek–Brown criterion. The critical sliding surface of the 3D slope was comprehensively analyzed based on the stress, strain, displacement, and velocity data of the grid model at the critical failure state. First, the potential sliding zone of the 3D slope was identified through combined stress and strain analysis. Furthermore, two innovative approaches for determining the critical sliding surface of the 3D mine slope model were proposed. One method is based on the displacement change rate, which involves slicing the slope at appropriate angles and analyzing the displacement field of each slice. The other method employs a clustering algorithm to classify the slope grids into stable and sliding groups, with the interface of the two groups representing the sliding surface. Finally, a thorough error analysis was conducted on the critical sliding surfaces obtained from both methods. The sliding surface identified by K-medoids clustering closely matches the one based on the displacement change rate with an acceptable height error (root mean square error = 2.76 m). The results indicate a significant similarity in the sliding surface positions obtained by both methods, confirming their feasibility and reliability. Compared to traditional methods, such as displacement contour line and shear strain increment thresholds, the two proposed methods can achieve more precise localization of sliding surfaces.

Edge-aware multi-attention network for small object detection from LiDAR-based 3D point cloud

FESSNet: Boosting few-shot point cloud semantic segmentation with Feature-Enhanced Self-Support Network

Autonomous productivity measurement NDT-SLAM-based algorithm for excavator earthwork operations

Imperceptible adversarial attacks for 3D point clouds using dimension features and Gaussian kernel perturbations

MIFSO-Net: A 3D point cloud registration network using multilevel contextual feature interaction and saliency-Guided overlap enhancement

TIH: Transformer in hyperbolic space for 3D point cloud semantic segmentation

Exploring transfer learning for individual tree species classification by cross-platform point cloud

Quality-controlled registration of urban MLS point clouds reducing drift effects by adaptive fragmentation

TTM: A concise yet effective surface reconstruction approach for tunnel point cloud from mobile mapping system

Low-rank tensor optimization with total variation regularization for point cloud denoising