标题： 基于NeRF的3D线索可视化支持数据驱动的航天器位姿估计 显示英文标题

标题： NeRF-based Visualization of 3D Cues Supporting Data-Driven Spacecraft Pose Estimation

摘要： 在轨操作需要估计追踪航天器与其目标之间的相对6D位姿，即位置和方向。 虽然已经开发了数据驱动的航天器位姿估计方法，但由于对其决策过程缺乏理解，这些方法在实际任务中的应用受到阻碍。 本文提出了一种方法，用于可视化给定位姿估计器所依赖的3D视觉线索。 为此，我们使用通过位姿估计网络反向传播的梯度训练了一个基于NeRF的图像生成器。 这使得生成器能够渲染航天器位姿估计网络所利用的主要3D特征。 实验表明，我们的方法能够恢复相关的3D线索。 此外，它们还提供了关于位姿估计网络监督与其对目标航天器的隐式表示之间关系的额外见解。 显示英文摘要

摘要： On-orbit operations require the estimation of the relative 6D pose, i.e., position and orientation, between a chaser spacecraft and its target. While data-driven spacecraft pose estimation methods have been developed, their adoption in real missions is hampered by the lack of understanding of their decision process. This paper presents a method to visualize the 3D visual cues on which a given pose estimator relies. For this purpose, we train a NeRF-based image generator using the gradients back-propagated through the pose estimation network. This enforces the generator to render the main 3D features exploited by the spacecraft pose estimation network. Experiments demonstrate that our method recovers the relevant 3D cues. Furthermore, they offer additional insights on the relationship between the pose estimation network supervision and its implicit representation of the target spacecraft.