Unseen Object Pose Estimation via Registration
Jun Wu, Yue Wang, Rong Xiong · 2021
Many current object pose estimation methods require precise and dense instance model, which significantly limits their application in practice. Some category level pose estimation methods deploy variational auto encoder to learn the regularity of similar objects, and estimate the pose of new objects with the knowledge learned from the same class. But they are still constrained to the learned category, and sensitive to the initial guess for searching the pose to minimize difference between query and synthesized image. In this paper, we propose a globally registration based pose estimation methods for unseen objects, consists of a correspondence prediction network and a differentiable weighted solver to filter outliers and enhance accuracy. Our method is designed to learn the implicit feature points correspondence between reconstructed models to register, omitting the demand of precise instance model. At inference time, the refenrence images and query image are reconstructed into full model and partial observation model, respectively. Then we seek for the correspondence between the two models and evaluate the confidence in the prediction. Last, we weightedly solve the pose in a closed form. Our method shows comparable performance with state of the art, and achieves a better accuracy effeciency balance.