An On-Orbit Star Tracker Benchmark for Resident Space Object Detection and Attitude Estimation

Suthakar, Vithurshan; Kunalakantha, Perushan; Lee, Regina S. K.; Sohn, Gunho

doi:10.1038/s41597-026-07736-9

DUST Dataset

An On-Orbit Star Tracker Benchmark for Resident Space Object Detection and Attitude Estimation

¹, ¹, ¹, ¹

¹Department of Earth and Space Science, York University, Toronto, Canada

Scientific Data, Article in Press, 2026

Paper DOI PDF Dataset Code (Zenodo) Citation

DUST provides real on-orbit, star-tracker-class imagery for benchmarking RSO detection, multi-object tracking, and attitude estimation under dense star fields, stray light, lens flare, motion jitter, and tiny-object conditions.

Abstract

On-orbit benchmark for SSA and navigation

Space situational awareness increasingly relies on optical observations to detect and track resident space objects and estimate spacecraft attitude. DUST addresses the shortage of public, real, on-orbit, wide-field star-tracker-class datasets by providing near-infrared imagery from the Fast Auroral Imager aboard CASSIOPE. The dataset includes astrometrically calibrated stars, manually verified RSO annotations, spacecraft ephemeris and attitude, and image-quality metrics. It supports algorithm development for RSO detection in dense star fields, multi-object tracking under orbital motion, and attitude estimation from star-tracker-class imagery.

Core idea: use a science imager with star-tracker-like characteristics as a dual-use benchmark for space-object detection and attitude-estimation research.

Dataset

Dataset Highlights

1,378

Near-infrared image frames

4,237

Verified RSO instances

160

Distinct RSO transits

22

Observation sessions

35,877

Detected stellar instances

26° FOV

FAI near-infrared wide-field imaging

Pipeline

Data Generation Pipeline

The workflow retrieves Level-1 FAI imagery and metadata, performs star detection and astrometric calibration, annotates RSOs through manual and semi-automated workflows, and extracts star/RSO patches for validation.

Visual Overview

Figures and Validation Products

Geographic and celestial coverage of DUST

Observation ground tracks and plate-solved stellar sky coverage across the 22 DUST sessions.

Stacked observations show star trails in white and RSO annotations in yellow across diverse backgrounds and geometries.

Validation plots summarize spatial priors, astrometric residuals, bounding-box statistics, feature embeddings, and ROC curves.

DUST background stability and illumination artifacts

Session-level background statistics and polar illumination maps characterize flare, stray light, and Milky Way contamination.

Benchmark

Supported Tasks and Technical Validation

Supported Tasks

RSO detection in dense stellar fields
Multi-object tracking across consecutive frames
Star identification and attitude estimation
Small-object and low-contrast detection studies
Background robustness and domain-shift analysis

Why It Is Challenging

RSOs are often only a few pixels wide.
Stars and RSOs appear nearly point-like in individual frames.
Motion is curved because of spacecraft attitude dynamics.
Frames include realistic flare, gradients, and stray-light artifacts.

Validation Summary

Metric	Reported value	Interpretation
Plate-solving success	84% overall; median 92%	Most sessions are reliably astrometrically calibrated.
Astrometric residuals	68% within 240.5 arcsec (0.63 px); 95% within 485.4 arcsec (1.27 px)	Sub-pixel at 68% containment; ~1.3 px at the 95% level.
Track completeness	98.4% average	RSO labels are temporally coherent across frames.
Tiny-object regime	~76.8% of RSOs have width ≤5 px and height ≤5 px	The benchmark targets extreme small-object detection.
Native feature separability	ROC-AUC 0.753	Single-frame static features provide only weak separation.

Files

Data Records

The dataset is distributed in open formats including PNG, CSV, TXT, FITS, and HDF5-derived metadata products. Session folders include image frames, star correspondence files, WCS solutions, background statistics, YOLO labels, MOT tracking labels, and thresholded images.

DUST/
├── SET_1/                         # manual annotation sessions
│   └── YYYY_MM_DD/
│       ├── Background/
│       ├── Corr/
│       ├── H5/
│       ├── Images/
│       ├── RSO_Annotations/
│       │   ├── YYYY_MM_DD_MOT_gt.txt   # session-level MOT file
│       │   └── YOLO/                   # per-frame YOLO label files
│       ├── Thresholded_Images/
│       └── WCS/
└── SET_2/                         # semi-automated annotation sessions
    └── same structure as SET_1

BibTeX

@article{suthakar2026dust,
  title   = {An On-Orbit Star Tracker Benchmark for Resident Space Object Detection and Attitude Estimation},
  author  = {Suthakar, Vithurshan and Kunalakantha, Perushan and Lee, Regina S. K. and Sohn, Gunho},
  journal = {Scientific Data},
  year    = {2026},
  doi     = {10.1038/s41597-026-07736-9},
  url     = {https://doi.org/10.1038/s41597-026-07736-9}
}

Dataset BibTeX

@dataset{suthakar2026dust_data,
  title     = {DUST: An On-Orbit Star-Tracker Benchmark for RSO Detection and Attitude Estimation},
  author    = {Suthakar, Vithurshan and Kunalakantha, Perushan and Lee, Regina S. K. and Sohn, Gunho},
  year      = {2026},
  publisher = {Zenodo},
  version   = {V2},
  doi       = {10.5281/zenodo.20255672},
  url       = {https://doi.org/10.5281/zenodo.20255672}
}