PC2Model: ISPRS benchmark on 3D point cloud to model registration

PC2Model: ISPRS Benchmark on 3D Point Cloud to Model Registration We present PC2Model, a publicly available benchmark dataset for 3D point cloud-to-model registration, developed under the leadership of ICWG II/Ib and funded by the ISPRS Scientific Initiatives 2025. The dataset is designed to support training and evaluation of both classical and deep learning-based registration methods. Detailed information about the dataset can be found in this paper: PC2Model: ISPRS benchmark on 3D point cloud to model registration , Maboudi, M., Harb, S., Ferrao, J., Khoshelham, K., Turkan, Y., & Mawas, K. Any furthur updates could be found on our GitHub page. Dataset Overview PC2Model comprises 137 samples spanning seven categories: Source Category Samples Simulated Mechanical objects 25 Simulated Furniture 25 Simulated Home décor 25 Simulated Houses 25 Simulated Vehicles 25 Simulated Indoor spaces 6 Real Indoor spaces 6 Each sample includes: A 3D model as an .obj file. A corresponding transformed point cloud as an .e57 file. The associated ground-truth transformation matrix (point cloud-to-model) as a .txt file. Dataset Design PC2Model consists of simulated data for six categories and one real-world category. Simulated point clouds were generated using Helios++, integrated into Blender as an add-on and configured to replicate a Leica ScanStation P40 terrestrial laser scanner, incorporating realistic scanning artifacts such as mixed pixels, noise, occlusions, and point density variations. The Blender add-on and scripts used to generate the dataset are available here. 3D models were sourced from publicly available datasets and repositories, including ABC, ModelNet40, Fusion 360 Gallery Dataset, Thingi10K, Sketchfab, and the ISPRS benchmark on indoor modelling (Khoshelham et al., 2020). Transformations Each point cloud was subjected to a rigid-body transformation within predefined bounds, serving as the ground truth for registration algorithms. The transformations include: Rotation: angles in [0 deg, 360 deg] around each axis. Translation: factor in [-5, 5] x point cloud extent per axis. Scaling: applied with 50% probability, factor in [0.5, 1.5]. Evaluation Metrics The following metrics were used for evaluating registration performance on the PC2Model benchmark, with ICP (via CloudCompare) as a baseline: LOA (Level of Accuracy): Mean and median closest-point distance between the registered point cloud and the model surface. LOC (Level of Coverage): Ratio of sampled model surface points that are covered by the point cloud within a predefined distance threshold. Transformation error: Includes normalized translation error (with respect to the bounding box diagonal) and rotation error computed as the geodesic distance between estimated and ground-truth rotations. The evaluation results are available in our paper. Any furthur updates could be found on our GitHub page.