A 6D Object Pose Estimation Network Using Zone-Coding Coordinates and Error-Eliminating Masks
Jian Zhang, Bo Liu, Nengqi Zhang, Yongpeng Tian · 2025
The two-stage method based on the establishment of 2D-3D dense correspondence and the Perspective-n-Point (PnP) algorithm is a solution to the 6 D object pose estimation problem. In this article, Zone-Coding Coordinates and Error-Eliminating Masks (ZE) are added to a previous network named Coordinates-Based Disentangled Pose Network (CDPN) to propose ZE-Pose. By testing ZE-Pose on LINEMOD and LINEMOD Occlusion datasets, its performance is evaluated. Compared to the rotation head of CDPN, ZE-Pose achieved an improvement of about$\text{2. 9 \%}$and 17.8 % on the accuracy metrics of$5^{\circ} 5 ~\text{cm}$and Average Distance of Descriptors (ADD) respectively. Moreover, the proposed method outperformed the complete CDPN network by about 1.8 % on both metrics while streamlining the network structure. In comparison with other related works, ZE-Pose also demonstrated competitive results.