Learning 3D Rotations from Point Cloud Data

This paper presents a deep learning (DL) based method for 3D rotation prediction on point cloud data. The proposed approach utilizes a single graph convolutional layer to capture meaningful geometric features and further regresses rotational quaternion via shallow multi-layer perceptron (MLP). Experimental results show that the method outperforms traditional registration algorithms, such as Iterative Closest Point (ICP), across multiple evaluation metrics, and also surpasses several DL-based approaches in various criteria. Moreover, the model is designed with only a few thousand parameters, making it highly lightweight and computationally efficient.