3D point cloud object classification with PointNet-Lite and data augmentation

We introduce PointNetLite, a lightweight variant of PointNet, designed for efficient 3D object classification in resource-constrained settings. By reducing architectural complexity, it achieves a model size that is less than a quarter of PointNet, with training speeds over twice as fast and inference times reduced by more than three-fold. PointNetLite incorporates innovative data augmentation techniques, including real-world scaling to preserve metric object dimensions, and rotation augmentation to improve robustness. These strategies enhance the model's resilience to affine transforms and geometric variance. Experimental results show that PointNetLite achieves better classification accuracy than PointNet, and performs competitively with PointNet++ and DGCNN, while maintaining superior efficiency.