Point-based networks have been widely used in the semantic segmentation of point clouds owing to the powerful 3D convolution neural network (CNN) baseline. Most of the current methods resort to intermediate regular representations for reorganizing the structure of point clouds for 3D CNN networks, but they may neglect the inherent contextual information. In our work, we focus on capturing discriminative features with the interactive attention mechanism and propose a novel method consisting of the regional simplified dual attention network and global graph convolution network. Firstly, we cluster homogeneous points into superpoints and construct a superpoint graph to effectively reduce the computation complexity and greatly maintain spatial topological relations among superpoints. Secondly, we integrate cross-position attention and cross-channel attention into a single head attention module and design a novel interactive attention gating (IAG)-based multilayer perceptron (MLP) network (IAG–MLP), which is utilized for the expansion of the receptive field and augmentation of discriminative features in local embeddings. Afterwards, the combination of stacked IAG–MLP blocks and the global graph convolution network, called IAGC, is proposed to learn high-dimensional local features in superpoints and progressively update these local embeddings with the recurrent neural network (RNN) network. Our proposed framework is evaluated on three indoor open benchmarks, and the 6-fold cross-validation results of the S3DIS dataset show that the local IAG–MLP network brings about 1% and 6.1% improvement in overall accuracy (OA) and mean class intersection-over-union (mIoU), respectively, compared with the PointNet local network. Furthermore, our IAGC network outperforms other CNN-based approaches in the ScanNet V2 dataset by at least 7.9% in mIoU. The experimental results indicate that the proposed method can better capture contextual information and achieve competitive overall performance in the semantic segmentation task.

2D-3D registration is increasingly being applied in various scientific and engineering scenarios. However, due to appearance differences and cross-modal discrepancies, it is demanding for image and point cloud registration methods to establish correspondences, making 2D-3D registration highly challenging. To handle these problems, we propose a novel and automatic solution for 2D-3D registration in Manhattan world based on line primitives, which we denote as VPPnL. Firstly, we derive the rotation matrix candidates by establishing the vanishing point coordinate system as the link of point cloud principal directions to camera coordinate system. Subsequently, the RANSAC algorithm, which accounts for the clustering of parallel lines, is employed in conjunction with the least-squares method for translation vectors estimation and optimization. Finally, a nonlinear least-squares graph optimization method is carried out to optimize the camera pose and realize the 2D-3D registration and point colorization. Experiments on synthetic data and real-world data illustrate that our proposed algorithm can address the problem of 2D-3D direct registration in the case of Manhattan scenes where images are limited and sparse.

Point cloud data acquired through LiDAR technology enables the rapid reconstruction of complex indoor building structures. However, due to the discrete nature of point clouds, they fail to accurately represent the geometric dimensions of building components and cannot be directly applied to digital model construction. Conventionally, this limitation necessitates manual modeling in specialized BIM software to integrate both geometric and semantic information of building structures, which is labor-intensive and time-consuming. To address this, we propose an automated method for geometric feature extraction and BIM reconstruction, enabling more efficient and accurate modeling processes. By segmenting building components and extracting their geometric features, the method automates the construction of building structural entities based on the IFC (Industry Foundation Class) standard, which is an open and vendor-neutral modeling standard widely used in the BIM domain.Specifically, the approach starts by filtering ceilings and floors using histograms of height values, as they geometrically represent planar structures and can be represented by footprints, which commonly are constructed from closed contours formed by projecting wall segments onto horizontal planes. To ensure accurate footprint representation, non-wall objects, such as furniture, are first excluded from the scene. For this purpose, a series of viewpoints are used to simulate camera positions and generate image sequences. Coupled with these image sequences, a pretrained large-scale language-image model, YOLO-World, is applied to identify the bounding boxes of the furniture, while the SAM2 model is used to segment individual entities. The segmented pixels are then back-projected and aggregated in 3D space to isolate and exclude non-building objects. Once the walls are identified, the point clouds are processed using a region-growing and merging algorithm to extract multiple facades, which are projected onto horizontal planes to generate line segments. Based on these line segments and the scene’s bounding box, the horizontal plane is divided into cell partitions, and an energy optimization-based graph-cut algorithm is applied to identify the optimal cell set, with the resulting closed contours representing the footprints. These footprints are then extruded along the height direction into 3D geometries to generate Ifcwall, Ifcceiling, and Ifcfloor objects through the Ifcopenshell library, producing a complete and standardized IFC model.This method was validated in complex indoor environments with furniture such as tables, chairs, and sofas, demonstrating high precision in reconstructing fundamental building components like walls, floors, and ceilings. By providing an automated and efficient solution for simple indoor structure reconstruction, the approach lays the groundwork for modeling more intricate scenarios and facilitates the development of intelligent, sustainable digital twin models to support comprehensive lifecycle building management.