Point Cloud Pretraining Dataset Effects on MaskPoint Classification Performance

Point cloud data, which represent 3D objects or environments, play a critical role in applications such as autonomous driving, robotics, and augmented reality. While most Self-Supervised Learning (SSL) methods for 3D point cloud classification rely on pretraining with ShapeNet, a largescale synthetic CAD dataset, it remains unclear whether this dataset yields optimal downstream performance. This paper conducts a systematic evaluation of pretraining dataset effects on the MaskPoint SSL model. ShapeNet, ModelNet40, and ScanObjectNN variations which are synthetic and real-world datasets are evaluated to see how they affect training efficiency and classification accuracy. While previous studies using the PointMAE method showed that synthetic datasets like ShapeNet achieved high accuracies, the results of this paper surprisingly reveal that pretraining on the ScanObjectNN PB_T50_RS variant yields the highest classification accuracy of 92.87% on ModelNet40 and 88.30% on ScanObjectNN OBJ_ONLY, outperforming ShapeNet with significantly less pretraining time (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3. 6}$</tex> hours against <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 1. 8 7}$</tex> hours). ScanObjectNN PB_T50_RS's lower size than ShapeNet (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 8 1 4}$</tex> seconds) causes its higher time per sample (1.135 seconds). The voting process is used to increase forecast resilience, hence increasing further accuracy. These results show that better generalization may be obtained from real-world datasets with noise and perturbations such as ScanObjectNN PB_T50_RS than from simulated CAD datasets. Expanding the size of such datasets could further enhance SSL model performance in point cloud classification tasks. These findings are especially relevant for industrial use cases such as digital twins, immersive training, and real-world monitoring where robust 3D understanding is critical.