Evaluating the Impact of Sampling Strategies on Classification Accuracy in Point-MAE for Point Cloud Classification
Résumé
Point-MAE is a recently introduced technique for deep learning on point cloud data, and prior research has demonstrated the significant impact of pretraining datasets on the classification downstream task performance. Another critical factor in working with point cloud data is the sampling method and the number of sampling points used. This paper explores different sampling techniques for point cloud data, with a focus on farthest point sampling, as employed in Point-MAE. In this study, Point-MAE is pretrained on two datasets using varying numbers of sampling points (512, 1024, and 2048). Each pretrained model is then fine-tuned with the same sampling densities on four datasets. Results indicate that higher sampling points generally improve classification accuracy. However, pretraining on ModelNet40 with 512 sampling points and finetuning on ScanObjectNN PB_T50_RS variant achieved a classification accuracy of 92.72%. These findings highlight the importance of selecting an optimal pretraining dataset and sampling density based on dataset complexity.