SO(3)‐Pose: SO(3)‐Equivariance Learning for 6D Object Pose Estimation

6D pose estimation of rigid objects from RGB-D images is crucial for object grasping and manipulation in robotics. Although RGB channels the depth (D) channel are often complementary, providing respectively appearance geometry information, it still non-trivial on how to fully benefit two cross-modal data. From simple yet new observation, when an rotates, its semantic label invariant while keypoint offset direction variant pose. To this end, we present SO(3)-Pose, a representation learning network explore SO(3)-equivariant SO(3)-invariant features estimation. The facilitate learn more distinctive representations segmenting with similar channels. communicate deduce (missed) detecting keypoints reflective surface channel. Unlike most existing methods, our SO(3)-Pose not only implements information communication between channels, but also naturally absorbs SO(3)-equivariance knowledge images, leading better learning. Comprehensive experiments show that method achieves state-of-the-art performance three benchmarks. Code available at https://github.com/phaoran9999/SO3-Pose.