TY - GEN
T1 - SCPose-MLite
T2 - 2025 IEEE International Conference on Real-Time Computing and Robotics, RCAR 2025
AU - Chen, Yuran
AU - Wu, Xuesong
AU - Fu, Kangjia
AU - Zhang, Qi
AU - Yu, Sunquan
AU - Zhong, Rui
AU - Sun, Xiucong
AU - Zhang, Xiang
AU - Yi, Teng
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - Accurate pose estimation of space targets is of great significance for conducting on-orbit rendezvous, space debris removal and other tasks.In this paper,we propose a lightweight deep learning model, SCPose-MLite, which is based on the deep learning framework of URSONet and adopts MobileNet-V2 as the backbone network.By integrating the mixed pooling module, SCPose-MLite further enhances the generalization ability of the model while maintaining high accuracy.The experimental results show that SCPose-MLite is close to URSONet in terms of pose estimation accuracy, and at the same time, it has improved the generalization factor by 2.32 times, which can better meet the actual needs of space target pose estimation.Furthermore, this paper attempts to run the lightweight model on different devices, demonstrating that the training optimization brought by the mixed-precision quantization strategy and high-computing-power devices may fail in the training of low-parameter networks.
AB - Accurate pose estimation of space targets is of great significance for conducting on-orbit rendezvous, space debris removal and other tasks.In this paper,we propose a lightweight deep learning model, SCPose-MLite, which is based on the deep learning framework of URSONet and adopts MobileNet-V2 as the backbone network.By integrating the mixed pooling module, SCPose-MLite further enhances the generalization ability of the model while maintaining high accuracy.The experimental results show that SCPose-MLite is close to URSONet in terms of pose estimation accuracy, and at the same time, it has improved the generalization factor by 2.32 times, which can better meet the actual needs of space target pose estimation.Furthermore, this paper attempts to run the lightweight model on different devices, demonstrating that the training optimization brought by the mixed-precision quantization strategy and high-computing-power devices may fail in the training of low-parameter networks.
UR - https://www.scopus.com/pages/publications/105016842685
U2 - 10.1109/RCAR65431.2025.11139482
DO - 10.1109/RCAR65431.2025.11139482
M3 - 会议稿件
AN - SCOPUS:105016842685
T3 - RCAR 2025 - IEEE International Conference on Real-Time Computing and Robotics
SP - 606
EP - 611
BT - RCAR 2025 - IEEE International Conference on Real-Time Computing and Robotics
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 1 June 2025 through 6 June 2025
ER -