Star-Fusion Presents Discrete Celestial Orientation Model

Per arXiv submission 2604.26582 (submitted 29 Apr 2026), authors May Hammad and one coauthor present Star-Fusion, a multi-modal transformer architecture for discrete celestial orientation estimation. The paper reframes attitude determination as a classification problem over a spherical partition obtained via spherical K-Means, rather than a direct regression on RA/Dec coordinates. The paper reports Top-1 accuracy of 93.4% and Top-3 accuracy of 97.8% on a synthetic Hipparcos-derived dataset, and an inference latency of 18.4 ms on commodity (COTS) hardware.

Per the paper, the model combines three complementary branches in a tripartite fusion strategy:

• •a SwinV2-Tiny transformer backbone for photometric feature extraction,
• •a convolutional heatmap branch for spatial grounding, and
• •a coordinate-based MLP for geometric anchoring. The authors use spherical K-Means to partition the celestial sphere into K topologically consistent regions to mitigate coordinate wrapping and boundary artifacts inherent to RA/Dec parameterizations. Experiments are reported on a synthetic dataset derived from the Hipparcos catalog; the submission includes measured inference latency and classification accuracies as primary evaluation metrics.

Editorial analysis: Techniques that convert non-Euclidean regression problems into discrete classification tasks recur in the literature when topology or periodic boundaries impair direct regression. Discrete partitioning combined with modern transformer-based encoders can reduce failure modes from coordinate wrapping, at the cost of discretization granularity and potential class-boundary sensitivity. For spacecraft and embedded applications, reported 18.4 ms inference latency on COTS hardware is directly relevant to onboard real-time requirements and power-constrained deployments, but real-flight validation remains the typical next step.

For practitioners: verify reported robustness on real star-tracker imagery and against sensor noise distributions not represented in synthetic Hipparcos-derived data; evaluate sensitivity to the chosen K in spherical K-Means and to class-boundary errors; benchmark against established LIS ("Lost-in-Space") pipelines and hybrid sensor-fusion stacks; and test latency and reliability across representative radiation-tolerant hardware and image pipelines.