A Real-Time 6-DoF Object Pose Estimation Method Based on Template Iteration

This paper presents a novel real-time 6-DoF object pose estimation framework leveraging template iteration combined with RGB information. By integrating object detection, template matching, and pose refinement techniques, the proposed approach constructs a pose estimation network characterized by strong generalization capabilities, real-time performance, and high accuracy. Traditional RGB-based 6-DoF pose estimation methods typically rely on computing the residual between rendered images of coarsely estimated poses and the input query images. However, due to the lack of sufficient 3D structural information, these methods often exhibit poor generalization to unseen objects. In this paper, we propose a 3D feature set-based pose iterative network that eliminates the need for rendering images on the input pose, thus enabling pose estimation without CAD models. Initially, object detection algorithms are employed to identify objects within the scene, followed by template matching to determine the initial pose. Subsequently, the pose iterative network is designed to refine the initial pose, achieving higher estimation accuracy. To ensure real-time performance, model pruning techniques are utilized to compress the pose iterative network, significantly reducing computational complexity while maintaining estimation accuracy. Experimental results demonstrate that our method exhibits robust generalization capability, real-time performance, and high accuracy across various scenarios, providing an efficient and reliable pose estimation solution for robotic vision and automation.