LoD2-Former Dataset: Multi-Modal Transformer-Based 3D Building Wireframe Reconstruction

This record contains two curated, multi-modal datasets for building-roof structure analysis and 3D roof reconstruction. Each dataset aligns, per building, a set of complementary 2D and 3D representations of the same roof — aerial RGB imagery, a rasterised height map, a 3D point cloud, and pixel-level roof-plane annotations — enabling research that fuses image-based and geometry-based cues. The datasets accompany LoD2-Former: Multi-Modal Transformer-Based 3D Building Wireframe Reconstruction (Abdelhedi, Panangian, Amrullah, Chaabouni-Chouayakh and Bittner, 2026), where they are used to train and evaluate end-to-end roof wireframe reconstruction from combined LiDAR and imagery. The data is distributed as two archives. 1. roofint.zip An aerial-imagery roof dataset of approximately 3,585 individual buildings, each identified by a tile-based basename. Aerial imagery — RGB roof crops originating from the mesh–image paired roof dataset of Ren et al. (2021), Intuitive and Efficient Roof Modeling for Reconstruction and Synthesis, ACM Transactions on Graphics 40(6). That dataset pairs aerial images with manually specified roof topology and optimised planar 3D polygonal roof meshes. Several processing stages are provided here: raw crops (images_original/), tightly-cropped unpadded crops (images_non_Padded/), padded and cropped variants (images_cropped/, images0/), and the final processed RGB tiles (images/). Ground truth — per-roof height maps (heightmaps/) and pixel-level roof-plane segmentation labels (annot/, with unprocessed versions in annot_raw/). Both are derived from the planar 3D roof meshes of Ren et al. by projecting the roof geometry into the image plane and cleaning the resulting plane regions; the vertex files supply roof corner coordinates in image space and the face files define the vertex sequences bounding each roof segment. 3D point clouds — plain-text X Y Z point clouds per roof (xyz/), sampled from the triangulated faces of the roof meshes at a consistent density of 10,000 points per building, following the dataset generation methodology of AIM2PC (Turki, Panangian, Chaabouni-Chouayakh and Bittner, ISPRS Geospatial Week 2025). Splits and tooling — predefined train/valid/test lists with spatial-split visualisations under splits/, together with the Python scripts used to build the dataset: proj.py (point cloud to 2D projection and roof-plane region cleaning), building_align.py and overlap.py (RGB to height-map alignment), and flip_images.py (orientation fixes). 2. Tallinn.zip A roof dataset for the city of Tallinn, Estonia (approximately 2,840 buildings), curated from the Building3D dataset (Wang, Huang and Yang, ICCV 2023): LiDAR point clouds — airborne LiDAR scans per building in xyz/ (X Y Z). Roof wireframe models — Building3D roof structure wireframes as Wavefront .obj files (obj/), encoding roof corner vertices and edges. To this we add the aligned 2D modalities produced for this work: per-building RGB roof images (images/), height maps (heightmaps/), and roof-plane annotations (annot/), along with train/valid/test splits and spatial-split visualisations (splits/). Splits are defined spatially rather than at random, so that no geographic area appears in more than one split. The old split/ folder additionally includes the source LoD2 CityGML city model (hooned_lod2-Tallinn.gml). The RGB imagery and the underlying national airborne laser scanning data both originate with the Estonian Land Board (Maa-amet, now the Estonian Land and Spatial Development Board), whose Geoportal distributes nationwide orthophoto coverage. Directory structure roofint/ images/, images_original/, images_non_Padded/, images_cropped/, images0/ — RGB roof crops at successive processing stages heightmaps/ — per-roof height maps annot/, annot_raw/ — roof-plane segmentation labels (processed / raw) xyz/ — 3D point clouds (X Y Z) train_list.txt, valid_list.txt, test_list.txt, splits/ — dataset splits and visualisations proj.py, building_align.py, overlap.py, flip_images.py — curation scripts Tallinn/ images/ — RGB roof images heightmaps/ — height maps annot/ — roof-plane labels xyz/ — Building3D airborne LiDAR point clouds obj/ — Building3D roof wireframe models (.obj) train_list.txt, valid_list.txt, test_list.txt, splits/ — splits and visualisations old split/ — source LoD2 CityGML model (hooned_lod2-Tallinn.gml) and earlier splits Per building, the aligned core modalities are: RGB image, height map, point cloud and roof-plane annotation — plus a roof wireframe for the Tallinn set. Files share a single building identifier across every folder, so a complete sample is assembled by looking up the same basename in each directory. License of this datasetThis dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).Users must provide attribution to both the original data provider and the authors of this dataset when using or redistributing it.