Robot-Based Auto-labeling System for 6D Pose Estimation

6D object pose estimation is an ongoing research area in the field of computer vision. Many existing methods rely on supervised deep learning models which require multiple accurate 6D pose annotations to predict object poses. However, labeling the 6D pose is complex and time-consuming in traditional methods. In this study, we propose a robotic-arm-based 6D object pose auto-labeling approach which has limited human interaction involved. Translations and rotations of the object in the camera coordinate system can be calculated using a sequence of known robot poses and the transformation between the camera and the robot. We also implemented our custom dataset generated by the auto-labeling system in the existing 6D object pose estimation approach. Evaluation results show that the model can recognize our own test dataset and attempted 90% accuracy using ADD metric with 0.05 threshold.