SDCNet: A Symmetry-Driven Network for Category-Level Object Pose Estimation

Object pose estimation aims to predict the position and rotation angle of a target object in 3D space, providing critical spatial pose data for precision measurement tasks, thus significantly improving the accuracy and stability of measurement results. However, for the category-level object pose estimation task, it is often difficult for traditional methods to achieve good generalization performance due to the multiple geometrical differences that may exist in similar objects. To address this technical difficulty, this study innovatively proposes a symmetry-driven category-level object pose estimation network, SDCNet. The core innovations of the framework are: first, two different deformation strategies are designed by analyzing the topological features of the objects, which are able to effectively characterize the morphological differences of the same kind of objects; second, a lightweight end-to-end network design is adopted to directly predict the spatial pose parameters of the objects Secondly, a lightweight end-to-end network design is adopted to directly predict the spatial position parameters of the objects, which significantly reduces the computational complexity while ensuring the accuracy. The experimental data show that SDCNet achieves 94.6% performance on the IoU<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> metric, 91.8% on the IoU<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">75</inf> metric, and 91.6% on the 10°5cm metric (CAMERA25).