Silhouette-Based 6D Object Pose Estimation

Xiao Cui, Nan Li, Chi Zhang, Qian Zhang, Wei Feng, Liang Wan

Abstract: For a long time, deep learning-based 6D object pose estimation networks have lacked the ability to address the problem of pose estimation of the unknown objects beyond the training datasets, due to the closed-set assumption and the expensive cost of high-quality annotation. Conversely, traditional methods struggle to achieve accurate pose estimation for texture-less objects. In this work, we propose a silhouette-based 6D object pose estimation method. being a conventional method As a traditional method, our approach achieves high accuracy without any need of annotation data, demonstrating excellent generalization. Additionally, we employ silhouette to mitigate texture dependency issues, ensuring effectiveness even in the case of textureless objects. In the method, we introduce a dimensionality reduction strategy for \(\textrm{SE}\left( 3\right) \) pose space, accompanied by theoretical proofs, which make it possible to perform pose estimation through search, rendering, and comparison in a reduced-dimensional space efficiently and accurately. Experimental results demonstrate the high precision and generalization of the proposed method. Our code is available at https://github.com/worldTester/STI-Pose.