SIGNN++ identifies stars in 98.5% of a real-camera test set

A blurred night-sky image can still carry enough geometry to identify its stars. SIGNN++, a new graph-based method, solved 98.5% of images in a real-camera random-view test, compared with 59.7% for the earlier SIGNN method and lower rates for two other baselines.[1]

The preprint, submitted to arXiv on 20 October 2026, adds more realistic training noise, transformed angular-distance features and robust alignment to the earlier system. It targets star identification and camera orientation for nighttime settings where satellite-navigation signals are unavailable.[1]

A constellation is more than a picture

Human observers recognise the Big Dipper because its arrangement is distinctive. An automated system faces a harder task: the image may contain only a small patch of sky, the camera can be tilted in any direction, and the stars' positions are affected by detection error. The algorithm must compare relative geometry to a catalogue, establish which observed points correspond to which real stars, then estimate the attitude—how the instrument is oriented in three-dimensional space.[1]

SIGNN++ represents stars and their angular relationships as a graph, so it does not depend entirely on a full photographic scene being clean. Its training strategy includes distortions closer to those seen by ordinary cameras from the ground. A robust alignment stage then attempts to reject inconsistent matches. These engineering changes may sound modest; together they address the gap between a neat synthetic star map and an actual sky viewed through moving air and a trembling lens.[1][2]

On the real random-view dataset, solved rates were 59.7% for SIGNN, 70.1% for ROSIA, 79.1% for Tetra and 98.5% for SIGNN++. The new method reported an attitude error of 0.203 degrees. Tetra’s value was 0.183 degrees, but the authors did not find that difference statistically significant and noted that the sets of solved frames differed. The earlier SIGNN remained faster.[1][2]

What happens near the horizon

Near the horizon, atmospheric refraction shifts a star’s apparent position. SIGNN++ retained a solved rate of roughly 77% in the relevant tests. That is where its more realistic noise treatment and geometric matching become useful: visible points need not land exactly where an idealised sky predicts.[1][2]

The researchers also used sequential and random-view real-camera datasets and illustrated difficult cases involving streaks and duplicate centroids caused by motion blur. In one example, a star's streak could be picked up as two points. The graph-based matcher could still recover the arrangement. Real-image evaluation strengthens a star-identification study substantially compared with synthetic-only demonstrations, but the dataset does not stand in for every latitude, season, lens, exposure setting or vehicle vibration.[1]

There is another crucial distinction: recognising star directions is not the same as independently determining where you are on Earth. A camera can estimate orientation from the catalogue and observation geometry, but a full navigation solution needs appropriate time information, calibration and potentially additional observations or sensors. The paper centres on identification and attitude, not proof of a standalone replacement for GNSS or inertial navigation in an operational vehicle.[1]

An old reference system with a new matcher

The system improves a specific link in celestial navigation: matching imperfect camera observations to known stars. Combining that capability with calibrated timing, instruments and other sensors is the separate work of building a reliable navigation system.[1]

Report an error or suggest a correction