6D Object Pose Estimation Enhanced by Structural Constraint

Aiming at the problem that the 6D object pose estimation method based on the voting strategy ignores the structural information between keypoints, a 6D object pose estimation method enhanced by structural constraint (SC-Pose) is proposed. This method defines a shape descriptor designed to describe the structured information between the 2D keypoints of the object. By enhancing the keypoint structural loss to constrain the predicted shape descriptor to be close to the ground-truth shape descriptor, the positioning of the 2D keypoints is more accurate, and thereby ultimately e incorporating the accuracy of 6D object pose estimation. Results on the LINEMOD, OCC-LINEMOD and TruncationLINEMOD datasets show that SC-Pose can significantly boost the accuracy of 6D object pose estimation.