Performance Analysis of Deep Learning Models for 3D Point Cloud Classification: A Comparative Study on the TogogBali3d Dataset

The classification of 3D objects from point cloud data is a fundamental task in computer vision, with significant advancements driven by deep learning models, such as PointNet and dynamic graph CNN (DGCNN). Although these models have shown high performance on standard benchmark datasets, such as ModelNet10and ScanObjectNN, their efficacy on culturally specific specialized datasets remains largely unexplored. This paper presents an experimental study to evaluate and compare the classification performance of PointNet and DGCNN using a novel dataset, TogogBali3d, which comprises eight classes of traditional Balinese statues (Togog). We benchmarked the performance of these models on the TogogBali3d dataset and compared the results with their established performance on ModelNet10 and ScanObjectNN. This study aims to assess the generalization capabilities of state-of-the-art 3D classification models and highlight the challenges associated with classifying objects with intricate and unique geometric features. The DGCNN model achieved an Overall Accuracy (OA) of 61.25%, narrowly outperforming PointNet's 60.83% OA. The performance drop compared to generic benchmarks confirms the challenge posed by the fine-grained geometry of the dataset. A detailed analysis shows that while PointNet's global features are highly effective for specific macro-level classes (such as Punakawan and Pengapit), DGCNN's local feature aggregation provides marginally better overall consistency, particularly in achieving superior recall for classes defined by localized detail (such as Penabuh and Dewi).