Abstract
Large-scale 2D datasets have been instrumental in advancing machine learning;however, progress in 3D vision tasks has been relatively slow. This disparityis largely due to the limited availability of 3D benchmarking datasets. Inparticular, creating real-world point cloud datasets for indoor scene semanticsegmentation presents considerable challenges, including data collection withinconfined spaces and the costly, often inaccurate process of per-point labelingto generate ground truths. While synthetic datasets address some of thesechallenges, they often fail to replicate real-world conditions, particularlythe occlusions that occur in point clouds collected from real environments.Existing 3D benchmarking datasets typically evaluate deep learning models underthe assumption that training and test data are independently and identicallydistributed (IID), which affects the models' usability for real-world pointcloud segmentation. To address these challenges, we introduce the BelHouse3Ddataset, a new synthetic point cloud dataset designed for 3D indoor scenesemantic segmentation. This dataset is constructed using real-world referencesfrom 32 houses in Belgium, ensuring that the synthetic data closely aligns withreal-world conditions. Additionally, we include a test set with data occlusionto simulate out-of-distribution (OOD) scenarios, reflecting the occlusionscommonly encountered in real-world point clouds. We evaluate popularpoint-based semantic segmentation methods using our OOD setting and present abenchmark. We believe that BelHouse3D and its OOD setting will advance researchin 3D point cloud semantic segmentation for indoor scenes, providing valuableinsights for the development of more generalizable models.