Project Description:
This project develops a constellation detection system with two parallel components: a synthetic data generation pipeline and a benchmarking study of object detection models. The synthetic pipeline uses International Astronomical Union (IAU) celestial coordinate data to create 2D dot-plot skeleton images of constellations, which are fed into a generative image model to generate realistic star field images with accurate geometry. These synthetic images are used to address the scarcity of labeled astronomical training data. Three object detection architectures ( YOLO26 (CNN-based), RF-DETR (transformer-based), and D-FINE (hybrid) ) are fine-tuned across multiple model sizes on real, synthetic, and combined datasets, with performance evaluated on a testing set of 74 real constellation sky images using mAP50 and mAP50-95 metrics. The real dataset consists of 1,750 annotated images spanning 14 IAU constellations (out of 88 official constellations), from an online source. The ultimate goal is a fine-tuned detection model suitable for embedded deployment (e.g. on a Raspberry Pi), alongside a reproducible generative model-based image generation pipeline generalizable to all 88 IAU constellations and other astronomical bodies.
