Conference Agenda

This contribution introduces GT-LOD3, a new benchmark dataset designed to advance semantic Level of Detail 3 (LOD3) building reconstruction from UAS-based photogrammetric point clouds. Existing benchmarks primarily focus on mesh- or point-level semantic labelling, façade segmentation, or LOD2-level modelling, but high-quality, geometry-accurate LOD3 ground truth paired with real-world photogrammetric observations are still limited. GT-LOD3 fills this gap by offering paired UAS point clouds and manually modeled LOD3 reference data in CityGML format, enabling research on window-level facade reasoning, geometric regularization, and instance-level shape recovery.

The benchmark currently consists of two subsets featuring different architectural styles and environmental conditions: (1) a urban block in Gold Coast (Lakewood, Ohio, USA), and (2) the Technical University of Munich (TUM) campus. The accompanying LOD3 reference models contain explicit window geometry, enabling detailed evaluation of both detection performance and polygon-level geometric accuracy.

We further provide a baseline reconstruction pipeline that combines point-cloud semantic segmentation, facade-aligned 2D projection, window region extraction, and geometric back-projection into CityGML. An evaluation protocol is presented including pixel-level metrics (IoU, precision, recall, F1) and instance-level detection metrics based on optimal assignment via the Hungarian algorithm.

Building wireframe reconstruction from LiDAR faces challenges due to sparse and incomplete point cloud data. We present LoD2-Former, a multi-modal Transformer architecture that fuses aerial LiDAR and optical imagery for end-to-end 3D roof wireframe reconstruction. Unlike existing point-cloud-only methods, our dual-backbone approach with bidirectional cross-modal attention leverages complementary geometric and visual information. Experiments on two datasets show consistent improvements in edge detection metrics, with edge F1-scores increasing from 0.874 to 0.899 on Tallinn and 0.968 to 0.974 on Roof-Intuitive, while substantially boosting corner recall (0.630 to 0.729) in complete-data settings. We also contribute a curated multi-modal subset of Building3D with aligned LiDAR and aerial imagery to facilitate future research.

In this paper, a method exploiting aerial LiDAR point clouds to build realistic building meshes suitable for electromagnetic simulation is proposed. One of the main challenges lies in reconstructing regularized building meshes with low polygonal density. Optimization-based methods, commonly used for building reconstruction from point clouds, are highly data-driven, making the quality of results dependent on the quality of input data. Aerial LiDAR scans can be incomplete or sparse, for instance due to occlusion. A novel LoD2 buildings reconstruction method based on deep learning is proposed, assuming that deep learning methods are more robust to incomplete or sparse data than optimization-based methods. A parametric building model is introduced, based on the Weighted Straight Skeleton algorithm, which generates realistic roofs from a building footprint and an associated set of slopes, and subsequently extrudes the roof to the specified building height. This parametric approach guarantees that a given set of parameters (height, footprint and slopes) produces a regularized building mesh with low polygonal density. A multimodal model, named Point2WSS, was trained to recover the variable number of building's continuous parameters from its corresponding point cloud. This approach enables the generation of realistic building meshes suitable for electromagnetic simulation, if the predicted parameters accurately approximate real-world values.

The modeling of buildings suffers from a dichotomy between generic and specific representations: the lack of domain knowledge in flexible models that can represent many shapes, and the restricted geometry of pre-specified parametric building primitives. To fill this gap, we propose using general boundary representations enriched with automatically recognized and enforced geometric constraints derived from human-made regularities. The proposed reasoning process relies on the statistics of the planar point groups extracted from airborne-captured point clouds. Hence, a chosen significance level is the only process parameter. To enforce the creation of sound solids, we apply manifold constraints for the generation of the boundary representations. The feasibility and usability of the approach are demonstrated by evaluating an airborne-captured laser scan containing approximately 7,600 buildings over an area of 50 km^2 featuring both inner-city and rural landscapes.

We present P3, a large-scale multimodal dataset for building vectorization, including aerial LiDAR point clouds, aerial images, and vectorized 2D building outlines, collected across three continents. P3 contains over 10 billion LiDAR points with decimeter-level accuracy and RGB images at a ground sampling distance of 25 centimeters. While many existing datasets focus on the image modality, P3 offers a complementary perspective by incorporating dense 3D information. We demonstrate that LiDAR point clouds serve as a robust modality for predicting building polygons, both in hybrid and end-to-end learning frameworks. Moreover, fusing LiDAR and imagery further improves accuracy and geometric quality of predicted polygons. The P3 dataset is publicly available, along with code and pretrained weights of three state-of-the-art models for building polygon prediction at https://github.com/raphaelsulzer/PixelsPointsPolygons .

Accurate building height estimation plays a crucial role in large-scale 3D urban reconstruction. However, conventional stereo matching approaches often suffer from mismatches around building edges, leading to unreliable height retrieval in dense urban areas. To address this issue, this paper presents a novel method for building height estimation based on contour vector registration integrated with the vertical line locus technique. The proposed framework first automatically matches building contour vectors extracted from stereo high-resolution satellite images. Then, for each paired contour, a range of candidate heights is searched using a rational function model to project the reference contour from the image space to object space and then reproject it onto the conjugate image. The elevation that maximizes the overlap ratio between projected and paired contours is identified as the optimal roof elevation. Building height is subsequently derived by subtracting the ground elevation from the estimated roof elevation. Experiments conducted on SuperView-1 (SV-1) satellite stereo images over Jiuyuan District, Baotou, Inner Mongolia, China, demonstrate the effectiveness of the proposed method. The resulting building height estimates achieve a root mean square error of 0.84 m compared to manual measurements, showing strong agreement (r = 0.9993). The proposed contour-based stereo registration approach provides a robust and efficient solution for building height extraction from high-resolution satellite data, supporting precise urban 3D modeling and large-scale spatial analysis.