PaperlayerView exact revision on arXiv

Source-linked AI summary

SIGNN++: Robust Graph Neural Star Identification on Earth

Jeongbin Sohn, Jina Lee, Max Yen, Dowan Gwon, Jeahn Han, Floyd Hepburn-Dickins, Brian Coltin, Ryan Soussan, Pyojin Kim

arXiv:2610.30186v1astro-ph.IMcs.LG

TL;DR

Real-terrestrial reliability of learning-based star identification remains unclear because prior evaluations focused mainly on synthetic or non-terrestrial settings. SIGNN++ extends SIGNN with terrestrial-noise training, Fourier edge features, and RANSAC-based Perspective Star Alignment; it reports improved robustness and solved rate in terrestrial evaluation, with current real-world experiments limited to constant-velocity camera motion.

  • Problem

    Learning-based SID methods have mainly been evaluated in synthetic or non-terrestrial settings, leaving their reliability on real terrestrial images unclear.

    Source: §I · p. 1

  • Method

    SIGNN++ combines terrestrial-noise training, Fourier-transformed edge features, and RANSAC-based Perspective Star Alignment for star identification and attitude refinement.

    Source: §I · p. 1

  • Results

    SIGNN++ shows greater robustness than other methods as light pollution increases and substantially improves solved rate over SIGNN, with PSA adding approximately 0.35 s on average.

    Source: D. Stellarium Dataset · p. 6

  • Takeaways & Limitations

    SIGNN++ remains robust to duplicated centroids caused by star streaks in the reported sequential images, and the conclusion reports more robust performance in real terrestrial environments than SIGNN and representative SID methods.

    Source: E. Real-World Dataset · p. 7, §V · p. 8

    Real-world experiments use constant-velocity camera motion, making it difficult to generalize current results to fully real-time scenarios.

    Source: F. Limitation · p. 8

Abstract

from arXiv · show

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/.

I. INTRODUCTION

Terrestrial star identification matters for GNSS-denied navigation, but its reliability on real ground-based images has received limited evaluation. SIGNN++ adapts a learning-based SID pipeline to terrestrial noise and evaluates it using real images and star-tracker attitude references.

I. INTRODUCTION

  • Only a few SID studies evaluate attitude estimation on real ground-based datasets, while terrestrial observations bring atmospheric refraction, cloud occlusion, and other noise.SIGNN also lacks real-world image validation and has limited robustness to combined noise.Source: §I · p. 1
  • SIGNN++ combines terrestrial-noise training data, Fourier-transformed edge-distance features, and joint refinement of star IDs and camera attitude.The edge-distance features are described as less sensitive to photometric variation.Source: §I · p. 1
  • The authors evaluate real ground-based night-sky images against a co-mounted Rocket Lab ST-RT star tracker and release timestamped images paired with tracker attitude measurements.Source: §I · pp. 1–2

II. RELATED WORK

Prior SID work includes deterministic geometric matching and learning-based methods, but evaluation on real terrestrial observations remains limited. SIGNN++ builds on SIGNN’s star-graph representation and adds a pipeline for terrestrial identification and attitude refinement.

II. RELATED WORK

  • Deterministic SID methods match geometric relations, while learning-based methods aim to improve noise robustness; prior evaluations and terrestrial systems leave real-image learning-based pipeline evaluation limited.Deterministic approaches include Grid, Pyramid, Log-polar, Tetra, and ROSIA,,,,; SIGNN is a graph-based learner, and prior terrestrial systems use deterministic initialization, external attitude systems, or tracking,,.Source: §II · p. 2

III. PROPOSED METHOD

  • SIGNN++ extracts star centroids, performs SIGNN-based identification, then uses Perspective Star Alignment to refine predicted IDs and camera attitude.Source: §III · p. 2

A. SIGNN

  • SIGNN represents detected stars as graph nodes and their geometric relationships as edges, giving the representation rotation invariance.Its two-layer Graph Attention Network aggregates direct- and 2-hop-neighbor information; relative magnitude is the node feature and angular distance is the edge feature.Source: A. SIGNN · p. 2

B. Fourier Transform in SIGNN

SIGNN++ transforms angular-distance edge features with Fourier features and uses Perspective Star Alignment to refine star IDs and attitude. PSA combines RANSAC-style inlier aggregation with correspondence-based nonlinear refinement.

B. Fourier Transform in SIGNN

  • Fourier features map one-dimensional angular distances into a higher-dimensional representation, increasing discrimination between geometrically similar edges.The sinusoidal mapping is intended to mitigate deep networks’ spectral bias,; this study sets hyperparameter L to 6 for best performance.Source: B. Fourier Transform in SIGNN · p. 2

C. Perspective Star Alignment (PSA)

  • PSA uses RANSAC-style sampling and catalog inliers to estimate rotation, then refines field of view and rotation through one-to-one correspondences and nonlinear least squares.It uses a 0.5° catalog-neighbor radius for candidate inliers; the reported PSA refinement achieves identification precision comparable to Tetra.Source: C. Perspective Star Alignment (PSA) · pp. 2–3

D. Image-Based Star Magnitude Estimation

The method estimates apparent stellar magnitudes from image photometry rather than assuming known magnitudes. It uses relative magnitudes for SIGNN, which is described as less sensitive to the reference-flux constant than ROSIA’s absolute-magnitude feature.

D. Image-Based Star Magnitude Estimation

  • Apparent magnitude is estimated by subtracting annulus-background contribution from aperture flux after image normalization, centroiding, and luminance conversion.Source: D. Image-Based Star Magnitude Estimation · p. 3
  • M = −2.5 log10(F) + C converts measured flux F to apparent magnitude M, with C determined by reference flux.The study empirically sets C to 12, while noting it may vary with sensor type and atmospheric conditions.Source: D. Image-Based Star Magnitude Estimation · p. 3
  • SIGNN uses relative magnitudes between neighboring stars and is therefore relatively robust to variations in C, whereas ROSIA uses absolute magnitude information and is more sensitive to it.Despite estimation errors, computed magnitude still provides useful node information for SIGNN.Source: D. Image-Based Star Magnitude Estimation · p. 3

E. Noise Application in the Simulated Dataset

SIGNN++ training uses a terrestrial-noise-aware simulation with mixed noise types and realistic star-loss and false-detection patterns. Evaluation spans synthetic noise ablations, Stellarium simulations, and real-world night-sky images.

E. Noise Application in the Simulated Dataset

  • Four noise types—false-star, drop, magnitude, and positional noise—are applied independently with probability 0.7, while 20% of frames remain noise-free.Noise-free frames provide a reference for learning the underlying geometric structure.Source: E. Noise Application in the Simulated… · pp. 3–4
  • Drop noise makes dimmer stars more likely to disappear, while false stars are more often dim and placed at random image locations.The false-star count ratio is sampled uniformly from [0, noise level].Source: E. Noise Application in the Simulated… · p. 4
  • Magnitude and positional noise use Gaussian perturbations with sampled standard deviations up to 1 and 0.01, respectively, and generate 14,000 training images per epoch.Each image is a random celestial-sphere view within a circular FOV cap of 25°.Source: E. Noise Application in the Simulated… · p. 4

IV. EVALUATION

  • Evaluation includes synthetic noise ablations, Stellarium terrestrial simulations, and real-world random-view and continuous night-sky images with star-tracker ground truth.For ROSIA, the angular inlier radius is set to 2 × 10−3 rad and the magnitude gate to 5.0 for real-image experiments.Source: §IV · p. 4

A. Dataset Generation

The evaluation datasets span catalog-based training and validation, controlled Stellarium scenes, and real terrestrial images. They include random-view images for identification comparisons and continuous sequences for real-time assessment.

A. Dataset Generation

  • The catalog dataset trains and validates SIGNN++ using stars brighter than magnitude 6 from BSC5.Samples use stars within a circular FOV with a 25° diameter, and the changing noise and viewing direction at each epoch make the effective training set virtually unlimited.Source: A. Dataset Generation · p. 4
  • The Stellarium dataset varies light pollution across Bortle levels 6–9 and separates sky elevations of 50°–90° from horizon elevations of 10°–30°.It contains 5,000 images, with 625 for each Bortle-level and horizontality combination; near the horizon, refraction introduces centroid positional errors.Source: A. Dataset Generation · pp. 4–5, Figure 5 · p. 4, Figure 4 · p. 4
  • The random-view real dataset contains iPhone 15 Pro Max images captured on three July dates, with clouds and motion blur present in many images.Images are 2000 × 2000 pixels with a 20° FOV and nominal 1.6 s exposure; Astrometry.net supplies reference star IDs and attitudes.Source: A. Dataset Generation · p. 5
  • The continuous real dataset comprises 15 sequences of eight to ten images, captured at 0.5 fps with a co-mounted iPhone and Rocket Lab ST-RT tracker.The tracker provides reference attitudes, but variable tracker latency prevents perfect synchronization with the camera.Source: A. Dataset Generation · p. 5
  • Because tracker-camera synchronization can mismatch, attitude errors are calibrated with the sequence’s median attitude and omitted when only one frame is identified.The sequence setup uses a rotation stage and a 500 ms iPhone exposure; the illustrated tracker centroid image is not an actual tracker image.Source: A. Dataset Generation · p. 5

B. Evaluation Metrics

Evaluation reports precision, recall, F1, solved rate, computation time, and attitude error. In catalog noise ablations, SIGNN++ ranks best overall, while Fourier-transformed edge features improve SIGNN’s robustness, especially to positional noise.

B. Evaluation Metrics

  • Precision, recall, and F1 score are reported alongside solved rate, computation time, and attitude error.Solved rate counts frames identifying at least three stars, or frames with attitude error below 1° when ground-truth IDs are unavailable; computation time excludes centroiding.Source: B. Evaluation Metrics · p. 5

C. Catalog Dataset

  • SIGNN++ ranks best overall, followed by Tetra, SIGNN+FT, ROSIA, and SIGNN in the catalog noise ablations.The ablations compare methods across 11 levels, including noise-free conditions; maximum levels are 50% drop noise, 50% false-star noise, 0.1 magnitude noise, and 0.01 rad positional noise.Source: C. Catalog Dataset · p. 5, Figure 7 · p. 5
  • Tetra is particularly vulnerable to positional noise, while increasing drop and false-star noise degrades its performance; magnitude noise leaves SIGNN and ROSIA nearly unchanged.Tetra does not use magnitude information, and SIGNN and ROSIA use accurate catalog magnitudes in this dataset.Source: C. Catalog Dataset · p. 6
  • Fourier transform improves SIGNN’s robustness across noise types, with a particularly noticeable gain under positional noise.The authors attribute this to learning higher-frequency components of the geometric representation.Source: C. Catalog Dataset · p. 6

D. Stellarium Dataset

In Stellarium tests, SIGNN++ leads in solved rate and F1 across Bortle levels and declines less at the most severe light pollution level. Near the horizon, it also maintains a substantially higher solved rate than the comparison methods.

D. Stellarium Dataset

  • SIGNN++ achieves the highest solved rate and F1 score at every Bortle level, with only a limited decline at level 9.The authors suggest its drop-noise training, which preferentially removes faint stars, helps mimic severe light-pollution star loss.Source: D. Stellarium Dataset · p. 6
  • SIGNN has the shortest computation time, while attitude-error differences between SIGNN++ and Tetra are below 0.001°.Tetra has the lowest attitude error at Bortle levels 6–8, and SIGNN++ at level 9.Source: D. Stellarium Dataset · p. 6
  • ROSIA is not best on any reported metric, and unreliable magnitude information increases its candidate search and computation cost.The authors note that its computation-time advantage becomes less significant in this evaluation.Source: D. Stellarium Dataset · p. 6
  • At the horizon, SIGNN++ reaches an approximately 77% solved rate while comparison methods remain relatively low, despite refraction-induced positional errors up to approximately 0.002 rad.The authors connect this robustness to SIGNN++’s tolerance of that level of positional noise in the ablation results.Source: Table II · p. 6, D. Stellarium Dataset · p. 6

E. Real-World Dataset

On real night-sky data, SIGNN++ performs strongly across random-view images and 15 continuous sequences, including challenging conditions such as cloud occlusion and star streaks. Its PSA refinement improves reprojection error, while computation-time results indicate practical—but not fully real-time—operation.

E. Real-World Dataset

  • On the random-view dataset, SIGNN++ achieved the best solved rate and F1 score, while Tetra had the lowest attitude error with no statistically significant difference from SIGNN++.Tetra computes attitude error only on solved frames, unlike SIGNN++, which includes more challenging unsolved frames where it still attempts estimation.Source: E. Real-World Dataset · pp. 6–7
  • PSA reduced random-view reprojection error from 8.11 ± 3.68 px to 4.77 ± 3.96 px by jointly optimizing FOV and R.The authors associate the remaining error likely with unmodeled camera distortion and centroiding errors, since they did not perform additional camera calibration.Source: E. Real-World Dataset · p. 7
  • SIGNN++ identified stars despite severe motion blur, cloud-induced false detections, and streaks detected as duplicate centroids.Its 1°–4° edge-connection range prevents connections between duplicated centroids in the described blur case.Source: E. Real-World Dataset · p. 7, Figure 9 · p. 7
  • SIGNN++ generally positioned within 0.3 s, whereas Tetra required more than 5 s on images where identification failed.Positioning occurs after image acquisition, so the authors do not characterize it as fully real-time.Source: E. Real-World Dataset · p. 7
  • Sequential images ran up to approximately 10 times faster than random-view images, likely because shorter exposures yielded fewer detected stars and reduced PSA processing time.The authors identify PSA as more computationally demanding than network inference and its reduced processing time as the main reason for the decrease.Source: E. Real-World Dataset · pp. 7–8
  • SIGNN++ achieved the highest solved rate in all 15 sequences, ranked first in attitude error in 11 and computation time in 14, and solved some frames where Tetra and ROSIA failed entirely.The authors note that the sequence count remains limited and more experiments are needed to validate generalization.Source: E. Real-World Dataset · pp. 7–8, Table IV · p. 8

F. Limitation

The real-world evaluation has limits in its assumed field of view, timing synchronization, and camera-motion conditions.

F. Limitation

  • SIGNN++ requires an approximate initial FOV; irregular star-tracker delay prevents exact synchronization, and constant-velocity camera motion limits generalization to fully real-time scenarios.The authors consider the approximate FOV assumption acceptable because FOV is determined by camera intrinsics.Source: F. Limitation · p. 8

V. CONCLUSION

SIGNN++ combines terrestrial noise-aware training, Fourier-transformed edge features, and RANSAC-based PSA for attitude refinement. Across simulated and real terrestrial datasets, it shows improved robustness to observation noise, light pollution, and challenging ground-based conditions compared with SIGNN and representative SID methods.

V. CONCLUSION

  • SIGNN++ combines a terrestrial noise model, Fourier-transformed edge features, and RANSAC-based PSA, with evaluations reporting improved robustness across simulated and real terrestrial conditions.The conclusion identifies real-time experiments on mobile platforms as future work.Source: §V · p. 8

Related papers

Loading 2610.30186v1…