Lightweight Real World 3-D Point Cloud Object Classification Using Synthetic Data

Existing methods for 3D point cloud object classification often rely on complex model architectures to learn refined local representations. Many of these methods also pre-train their models on large-scale datasets that contain a wide variety of instances or modalities. These approaches tend to achieve good accuracy at the cost of increased complexity, limiting their applicability in resource-constrained environments. In addition, the scarcity of real-world point cloud data has given rise to the Sim2Real learning paradigm, where models are trained on widely accessible synthetic datasets, and then applied to realworld datasets. However, this learning approach has experienced a significantly challenging 3D Domain Generalization (3DDG) problem. In this paper, we first propose a new PointNetLite model, a lightweight variant of the popular PointNet model, which can improve training and inference efficiency while maintain highly competitive performance. We further identify four factors contributing to the 3DDG problem, including view and point density differences between synthetic and real point clouds, inter-class scale differences, and intra-class rotational variability among real data. We then introduce three data augmentation techniques to address these problems, including a sampling method that simulates real-world scanning of 3D objects, a scale-preserving design for maintaining true object scales during data processing, and a rotation augmentation to enhance model robustness to object orientation variations. Our experimental results demonstrate that our framework is time and resource efficient while delivering strong performance. Remarkably, our approach exceeds the top classification accuracy on the Sim2Real benchmark by up to 8 %, offering a highly effective solution for real-time 3D object classification.