Evaluation of Point-MAE for Robust Point Cloud Classification Across Diverse Datasets

Several techniques have been proposed to classify point clouds. However, these techniques have not been validated over a wide range of datasets. Typically, the results obtained are specific to tests performed on a single dataset, limiting their generalizability and robustness. A technique called Point-MAE was recently introduced, exhibiting promising performance across multiple datasets. The objective of this paper is to thoroughly evaluate the Point-MAE technique by testing it across three different datasets for pre-training and performing classification on four diverse datasets. The results demonstrate that Point-MAE consistently delivers high performance, achieving a classification accuracy of up to 92% across varied datasets. This indicates its robustness and potential applicability across different scenarios. However, it was observed that training times vary significantly with dataset size, suggesting a need for further optimization to handle large-scale data more efficiently.