Self-supervised 6D object pose tracking using Rao-Blackwell particle filter and vision transformer

In the realm of computer vision, tracking the 6D pose of objects is crucial for ro-botic manipulation and navigation tasks. This paper introduces a self-supervised method for accurate 6D object pose tracking by integrating EfficientDet and a vi-sion transformer-based masked auto-encoder within the Rao-Blackwellized parti-cle filter framework. Addressing challenges in capturing pose uncertainties and encoding depth information, our approach discretizes 3D translations and rota-tions to estimate full posteriors over 6D poses. Validation on the YCB-Video dataset using the ADD(-S) metric demonstrates significant improvements, achiev-ing average accuracies of 72.80% with 50 particles and 75.51% with 200 particles, surpassing state-of-the-art methods. This research provides a robust framework for enhancing 6D object pose tracking in robotic applications.