Symmetric Ambiguity Spacecraft Dataset Plus (SASD+)

Symmetric Ambiguity Spacecraft Dataset Plus (SASD+) Overview Symmetric Ambiguity Spacecraft Dataset Plus (SASD+) is a large-scale synthetic visual dataset designed for monocular 6-DoF pose estimation and symmetric ambiguity analysis of non-cooperative spacecraft. Built upon the basic framework of the original SASD, SASD+ further enriches spacecraft types, observation scenarios, ambiguous pose samples, and background diversity, forming a more comprehensive and challenging benchmark for on-orbit visual perception research. Compared with existing mainstream spacecraft pose datasets such as SPEED , SASD+ provides richer, more standardized, and ambiguity-oriented supervision information, including: Dense 2D semantic keypoint coordinates Fine-grained keypoint visibility labels Geometry-consistent 6-DoF pose annotations Symmetric pair identity labels for ambiguity matching analysis SASD+ is particularly suitable for keypoint detection, visibility prediction, learnable PnP networks, symmetric ambiguity elimination, and robust monocular spacecraft pose estimation under complex on-orbit conditions. Camera & Pose Sampling Strategy SASD+ adopts a hybrid sampling pipeline, including spherical viewpoint sampling, ambiguity-oriented symmetric pose generation, and model-adaptive distance sampling, to fully cover general and ambiguous on-orbit observation scenarios. Angle Sampling Ranges Azimuth angle: [0°, 360°] Elevation angle: [-90°, 90°] Model-adaptive Distance Sampling Ranges To adapt to different spacecraft sizes, independent camera-to-target distance ranges are defined for each model: Fermi-GLAST: 200 m – 500 m ACRIMSAT: 30 m – 80 m CubeSat-1U: 30 m – 100 m Ambiguity-Oriented Pose Generation Different from purely random sampling, SASD+ explicitly constructs symmetric ambiguous poses. For each random initial pose, a 180° rotation around the spacecraft Z-axis is performed to generate geometrically symmetric but physically distinct poses. Such samples produce highly similar keypoint distributions in images, which effectively simulates the structural symmetry ambiguity widely existing in real spacecraft visual perception. The overall ratio of random normal poses to symmetric ambiguous poses is strictly fixed at 2:1. Background Configuration To simulate real on-orbit environments, the dataset contains 80% deep-space black backgrounds and 20% Earth backgrounds, ensuring sufficient scene diversity. Keypoint Definition Semantic 3D keypoints are predefined on each spacecraft model in the body coordinate system, ensuring cross-sample spatial consistency: Distributed on spacecraft main body and solar panels Contains a large number of structurally symmetric keypoints for ambiguity analysis Accurate visibility annotation for each keypoint under different viewpoints Different models have tailored keypoint layouts (10 or 16 keypoints) according to their structural complexity. Dataset Statistics Total images: 54,000 Resolution: 1920 × 1200 Train / Val / Test ratio: 10:1:1 Total models: 3 (Fermi-GLAST, ACRIMSAT, CubeSat-1U) Random : Symmetric pose ratio: 2:1 Black : Earth background ratio: 8:2 Camera Parameters The camera intrinsic settings are consistent with the classic SPEED dataset, ensuring fair comparison with mainstream pose estimation algorithms: Parameter Description Value Nu Image width (pixels) 1920 Nv Image height (pixels) 1200 fx Horizontal focal length 0.0172 m fy Vertical focal length 0.0172 m Px Horizontal pixel size 5.86 µm Py Vertical pixel size 5.86 µm