SIGNN++: Robust Graph Neural Star Identification on Earth
Authors: Jeongbin Sohn, Jina Lee, Max Yen, Dowan Gwon, Jeahn Han, Floyd Hepburn-Dickins, Brian Coltin, Ryan Soussan, Pyojin Kim
Organizations: Department of Mechanical and Robotics Engineering, Gwangju Institute of Science and Technology (GIST), Gwangju 61005, South Korea. · NASA Ames Research Center, CA 94043, USA.
We present SIGNN++, a learning-based star identification (SID) framework for celestial navigation in GNSS-denied nighttime environments. Although researchers have developed learning-based SID methods to improve robustness to observation noise, prior evaluations have focused mainly on synthetic or non-terrestrial environments, leaving their reliability under real terrestrial conditions unclear. To address these limitations, SIGNN++ extends SIGNN, a graph neural network-based SID method, with three components: a more realistic noise model for training data, Fourier Transform (FT) of angular-distance edge features, and RANSAC-based Perspective Star Alignment for identification and attitude refinement. The results show that SIGNN++ is more robust to terrestrial observation disturbances and achieves higher star identification accuracy and solved rates than the original SIGNN and publicly available SID baselines, while maintaining comparable attitude accuracy and practical computation time for continuous operation. The implementation code of SIGNN++ and datasets will be available at https://signnpp.github.io/signnpp/.
Figures & tables
Fig. 1: Celestial Navigation Overview and SIGNN++ Inference Results. The system calculates the camera attitude after it correctly identifies the stars in an image. The three images below show, in order, severe motion blur, cloud occlusion, and star streaks. SIGNN++ successfully identifies stars even in terrestrial images with severe noise conditions where conventional SID methods fail.
Fig. 2: SIGNN++ Architecture. It inputs a night sky image and outputs star IDs and the camera’s attitude. We use the center-of-gravity (CoG) method for star centroiding. Since centroiding is not the focus of this work, we use the CoG method as a simple and widely adopted baseline and provide identical centroid inputs to all SID methods for fair comparison. III-B and III-C describe the FT module and Perspective Star Alignment, respectively.
Fig. 3: Visualization of four types of training noise. Drop and false-star noise remove and add stars with fixed probabilities, respectively, while magnitude and positional noise apply perturbations sampled from Gaussian distributions.
Fig. 4: Visualization of four levels of light pollution rendered by Stellarium. The left image shows the Orion constellation observed without atmospheric visualization, while the right images show star centroiding results under atmospheric effects with different Bortle levels for the same scene. As the Bortle level increases, fewer stars are detected.
Fig. 5: Visualization of atmospheric refraction rendered by Stellarium. Near the horizon, atmospheric refraction introduces significant positional errors in star centroids.
Fig. 6: Experimental setup for acquiring sequential real sky images. The setup consists of a rotation stage, an iPhone, and a star tracker (ST). The iPhone image on the right is an actual image captured with a 500 ms exposure, whereas the ST centroid image is an illustration showing the expected centroid locations of stars rather than an actual ST image.
Fig. 7: Ablation on the four noises on the catalog dataset. The figure compares SIGNN++ with SIGNN [ 1 ] , Tetra [ 8 ] , and ROSIA [ 9 ] . Each ablation consists of 11 noise levels, including the noise-free case. The maximum noise levels are 50% for drop noise, 50% for false-star noise, 0.1 for magnitude noise, and 0.01 rad for positional noise.
Method
Metric
Bortle 6
Bortle 7
Bortle 8
Bortle 9
SIGNN
Solved Rate (%)
84.16
85.28
57.78
19.72
F1 (%)
50.37
51.90
38.85
32.79
Comp. Time (s)
0.01
0.01
0.01
0.01
SIGNN++
Solved Rate (%)
96.88
95.68
86.36
64.59
F1 (%)
98.94
98.98
99.06
98.69
Comp. Time (s)
0.29
0.29
0.14
0.05
TABLE I: Performance comparison across Bortle levels.
Method
Metric
Sky
Horizon
SIGNN
Solved Rate (%)
76.00
47.87
F1 (%)
51.140±18.444
39.921±15.588
Comp. Time (s)
0.012±0.003
0.011±0.003
SIGNN++
Solved Rate (%)
94.87
77.06
F1 (%)
99.065±2.072
98.773±2.817
Comp. Time (s)
0.271±0.263
0.125±0.201
TABLE II: Performance comparison according to horizontality.
Method
Solved Rate (%)
F1 (%)
Time Solved (s)
Time Failed (s)
Attitude Error (deg)
SIGNN
59.7
32.63
0.012
0.012
–
SIGNN++
98.5
98.32
0.366
0.001
0.203
ROSIA
70.1
69.15
0.532
0.696
0.328
Tetra
79.1
91.83
0.593
6.053
0.183
TABLE III: Performance comparison on the real random-view dataset.
Fig. 8: Inference results on random-view real night-sky images. The image contains overall motion blur caused by camera shake. SIGNN++ successfully solves the image. Since SIGNN++ connects edges only between stars separated by 1∘ – 4∘ , it is robust to duplicated centroids caused by motion blur and can successfully identify the stars.
Fig. 9: Inference results on sequential real night-sky images. SIGNN++ successfully solves six frames of the fifth sequence in Table IV . As shown in the cropped image on the right, a star streak is detected as two separate centroids, causing the other SID methods to fail identification. In contrast, SIGNN++ successfully identifies the stars despite this duplicated centroid caused by the streak.
Seq.
Setup
Method
Solved Rate (%)
Att. Error (deg)
Time Solved (s)
Time Failed (s)
1
90∘ , roll, 1∘ /s
SIGNN++
100.000
0.137
0.071
–
Tetra
100.000
0.151
0.428
–
ROSIA
88.889
0.222
0.568
1.698
2
90∘ , roll, 2∘ /s
SIGNN++
66.667
0.206
0.040
0.014
Tetra
33.333
0.404
0.755
4.477
ROSIA
33.333
0.483
0.356
0.636
TABLE IV: Performance comparison on sequential real images.