A Correspondence Selection Method for Enhancing Robustness in 6D Pose Estimation

초록

6D pose estimation is an essential technology that enables robots to accurately recognize and manipulate objects. Many deep learning approaches for 6D pose estimation are based on supervised learning techniques and they have replaced traditional methods with outstanding performance in various public datasets. However, the main shortcoming of supervised learning is the dependence on the training dataset. Network performance often deteriorates when observation conditions differ from those seen during training, such as sensor noise, corruptions, or perturbations. In this paper, we introduce a strategy for the trained network to produce consistent results in noisy environments. We train a correspondence search network with only the target’s 3D model and apply it to the real scene to get the initial candidates of correspondence. The best candidate is determined based on the initial candidates, matching potentials, and confidences of the candidates using the Bayesian process. Experimental results show that our method is robust to noisy and perturbed conditions within the evaluated object set and achieves more stable performance than other state-of-the-art methods. © 2004-2012 IEEE.

키워드