Next Generation Spacecraft Pose Estimation Dataset (SPEED+)

SPEED+ is the next-generation dataset for spacecraft pose estimation with specific emphasis on the robustness of Machine Learning (ML) models across the domain gap. Similar to its predecessor, SPEED+ consists of images of the Tango spacecraft from the PRISMA mission. SPEED+ consists of three different domains of imageries from two distinct sources. The first source is the OpenGL-based Optical Stimulator camera emulator software of Stanford’s Space Rendezvous Laboratory (SLAB), which is used to create the synthetic domain comprising 59,960 synthetic images. The labeled synthetic domain is split into 80:20 train/validation sets and is intended to be the main source of training of an ML model. The second source is the Testbed for Rendezvous and Optical Navigation (TRON) facility at SLAB, which is used to generate two simulated Hardware-In-the-Loop (HIL) domains with different sources of illumination: lightbox and sunlamp. Specifically, these two domains are constructed using realistic illumination conditions using lightboxes with diffuser plates for albedo simulation and a sun lamp to mimic direct high-intensity homogeneous light from the Sun. Compared to synthetic imagery, they capture corner cases, stray lights, shadowing, and visual effects in general which are not easy to obtain through computer graphics. The lightbox and sunlamp domains are unlabeled and thus intended mainly for testing, representing a typical scenario in developing a spaceborne ML model in which the labeled images from the target space domain are not available prior to deployment. SPEED+ is made publicly available to the aerospace community and beyond as part of the second international Satellite Pose Estimation Competition (SPEC2021) co-hosted by SLAB and the Advanced Concepts Team (ACT) of the European Space Agency.

The construction of the TRON testbed was partly funded by the U.S. Air Force Office of Scientific Research (AFOSR) through the Defense University Research Instrumentation Program (DURIP) contract FA9550-18-1-0492, titled High-Fidelity Verification and Validation of Spaceborne Vision-Based Navigation. The SPEED+ dataset is created using the TRON testbed by SLAB at Stanford University. The post-processing of the raw images is reviewed by ACT to meet the quality requirement of SPEC2021.

SPEED+ 是面向航天器姿态估计的下一代数据集，其核心聚焦于机器学习（Machine Learning, ML）模型在域偏移下的鲁棒性。与前代数据集一致，SPEED+ 包含来自PRISMA任务的探戈航天器图像。SPEED+ 包含两类不同来源的三种成像域。第一类来源为斯坦福空间交会实验室（Space Rendezvous Laboratory, SLAB）的基于OpenGL的光学模拟器相机仿真软件，该软件用于生成包含59960张合成图像的合成域。带标注的合成域按照80:20的比例划分为训练集与验证集，作为机器学习模型训练的核心数据源。第二类来源为SLAB的交会与光学导航测试台（Testbed for Rendezvous and Optical Navigation, TRON），该设施用于生成两个搭载不同照明源的硬件在环（Hardware-In-the-Loop, HIL）域：灯箱域与太阳灯域。具体而言，这两个域通过真实照明条件构建：使用带有漫射板的灯箱模拟反照率，使用太阳灯模拟来自太阳的高强度直射均匀光。相较于合成图像，它们能够捕捉到计算机图形学难以生成的极端场景、杂散光、阴影效应及各类通用视觉效应。灯箱域与太阳灯域均未标注，因此主要用于测试，这契合了天基机器学习模型开发中的典型场景：即在部署前无法获取目标太空域的标注图像。SPEED+ 数据集已面向航空航天及相关领域公众开放，作为由SLAB与欧洲空间局（European Space Agency）先进概念团队（Advanced Concepts Team, ACT）联合主办的第二届国际航天器姿态估计竞赛（Satellite Pose Estimation Competition 2021, SPEC2021）的配套数据集发布。

TRON测试台的构建部分由美国空军科学研究办公室（Air Force Office of Scientific Research, AFOSR）通过国防大学研究仪器项目（Defense University Research Instrumentation Program, DURIP）合同FA9550-18-1-0492资助，该项目标题为“天基视觉导航的高保真验证与确认”。SPEED+ 数据集由斯坦福大学SLAB利用TRON测试台构建。原始图像的后处理工作由ACT审核，以满足SPEC2021的质量要求。

搜集汇总

背景与挑战

背景概述

SPEED+是用于航天器姿态估计的下一代数据集，包含合成和硬件在环仿真图像，特别关注机器学习模型的跨领域鲁棒性。数据集分为有标签的训练/验证集和无标签的测试集，模拟了太空应用中目标领域标签不可用的实际场景。