A novel metric for 6D pose estimation

Abstract Current state-of-the-art evaluation methods for 6D pose estimation have several significant drawbacks. Existing error metrics can produce near-zero errors for poor pose estimations and are heavily dependent on the object point cloud used, resulting in vastly different outcomes for different objects. Furthermore, false detections are not considered at all. Evaluating pose estimators is crucial as it directly impacts the reliability and effectiveness of applications in robotics, augmented reality, and object manipulation tasks. Accurate evaluation ensures that pose estimators can be trusted to perform well in real-world scenarios, leading to better system performance and user satisfaction. In this paper, we conduct experiments to provide insights into how these metrics behave under isolated errors. We also introduce a novel evaluation approach with a metric independent of point clouds, making it applicable to a broader range of use cases than current metrics.