Rotation-invariant 3D convolutional neural networks for 6D object pose estimation

6D object pose estimation, crucial for applications such as scene understanding, AR/VR, and robotic grasping, focuses on determining an object's rotation and translation from single-view images. Despite advancements in 3D deep learning, existing methods still struggle with large shape variations, high training demands, and unseen poses. This paper addresses these issues by introducing a rotation-invariant neural network. We propose a rotation-invariant 3D convolutional network that processes a point cloud to predict per-point canonical coordinates, from which the 6D pose, is estimated. The network utilises relative distances and angles within a representative point set. Experiments on a public dataset show that our method outperforms several state-of-the-art baselines, excelling in handling novel poses and severe occlusions. Ablation studies further highlight the importance of the individual components.