G-SAM: A Robust One-Shot Keypoint Detection Framework for PnP Based Robot Pose Estimation

Robot pose estimation plays a fundamental role in various applications involving service and industrial robots. Among the methods used for robot pose estimation from a single image, the Perspective-n-Point (PnP) based approach is widely used due to its popularity and efficiency. An important part of this framework is keypoint detection. However, the current keypoint detection module used for PnP has two problems: Small number of input keypoints and Large error of input keypoints. This paper proposes a Grouping and Soft-ArgMax (G-SAM) framework to address these two problems: First, a simple and powerful Soft-ArgMax module followed by point subset selection is designed to address the problem of small number of input keypoints; Second, a grouping module is introduced, taking into account the texture and spatial structure information of the robot, to solve the problem of large error of input keypoints. Extensive experiments compare our proposed framework with existing state-of-the-art methods on several public datasets and demonstrate that it can provide more reliable, accurate and faster pose estimation for robotic applications.