SE(3)-Equivariance Learning for Category-Level Object Pose Estimation

Directly regressing 6-D poses and 3-D metric sizes from point clouds is an efficient solution for category-level object pose estimation tasks in the community of computer vision, which promotes the application of vision-based measurement in robotics. However, traditional regression-based methods exhibit a preference for specific orientations constrained in training data, which poses challenges in generalizing their inference to all potential rotations. To address this, we propose a novel architecture for category-level object pose estimation, introducing SE(3)-equivariance learning to enhance performance for objects in arbitrary poses. First, we utilize vector neurons to extract pose-sensitive features to derive SE(3)-equivariance between input point clouds and output 6-D poses. Second, we introduce a shape-aware pose regression strategy that incorporates invariant shape embeddings to constrain the flow of orientation information, enabling adaptation to shape variations for category-level pose estimation. Finally, to further improve performance, we introduce an implicit geometric boundary reconstruction scheme that enhances the learning of the global geometric structure of the object. The proposed method achieves well-generalized pose estimation in the full SE(3) space without an exhaustive data augmentation for all possible rotations, resulting in superior performance compared to state-of-the-art methods when generalizing to arbitrary object poses. Our experiments show that the proposed approach produces a significant advantage in the accuracy of pose estimation when compared with other regression-based methods in robotic manipulation tasks. Our code is available at https://github.com/hongzhidu/SE3-Category-level-Object-Pose.git