"Make 6D Pose Estimation FAST Again": Real-Time 6D Object Pose Estimation

In applications that require real-time feedback, such as quality control and robotic manipulation in industrial contexts, high speed and accuracy are still critical requirements [16], despite technological advances that have improved attitude estimation accuracy and speed. Systems in these scenarios must be able to analyze and respond to changes in the environment quickly and accurately. As deep learning methods and computing hardware develop rapidly, pose estimation technology has evolved, but finding the right balance between computational efficiency and pose estimation accuracy remains a major challenge, especially in dynamic and real-time applications. Currently, almost all of the estimation algorithms are not scalable enough in terms of estimation time and SOTA algorithm too slow in terms of estimation speed [31]. The aim of this study is to develop a fast pose estimator that can match or surpass current baselines in terms of accuracy and robustness, addressing the crucial trade-off between computation efficiency and pose estimation accuracy in real-time applications. The SOTA method GDRNPP is the basis of our work, and we optimize code and modify model structure, resulting in five model configurations: each one meeting the different estimation needs for estimation accuracy and time. The results of our study have been validated using four datasets, including the LM-O [51], YCB-V [51], T-LESS [14] and ITODD [21] dataset.