A Dataset for Dense Star Pattern Recognition in Wide-Field Cameras

Star Pattern Recognition Dataset of NAGSN Overview This repository contains the data used to develop and evaluate the Neighborhood-Aware Graph Siamese Network (NAGSN) for dense star pattern recognition in wide-field cameras. It includes semi-synthetic star fields, pre-generated training subsets, real-world CAPD-derived evaluation data, and simulated and real-camera distortion calibration data. The semi-synthetic benchmark is based on 180 sky regions sampled across the celestial sphere. Ninety regions are used for model development and the other 90 are held out for final evaluation. The benchmark includes four outlier ratios (5%, 10%, 15%, and 20%) and three distortion settings. The real-data benchmark contains 60 observational fields derived from the China Astronomical Plates Data (CAPD); these fields are used only for evaluation and are not used for model training. Files Path Description sim.tar.gz Semi-synthetic star-field data, organized by data split, outlier ratio, and distortion level. match_subsets.tar.gz Pre-generated paired star subsets used to train and validate the NAGS block. capd.tar.gz Derived tabular representations of 60 CAPD observational fields. Level 0 is distortion-free; levels 4-6 contain three calibrated real-camera distortion settings. Original CAPD FITS images are not redistributed in this archive. distortion_data.tar.gz Coordinate pairs and metadata used for simulated-distortion training and real-camera distortion adaptation/evaluation. Archive layout A typical processed sample has the following structure: <dataset>/<group>/<sample>/|-- img.csv|-- ast.csv`-- info.ini Common group names are: 0: distortion-free data. 1-3: simulated distortion levels. 4-6: distortion fields derived from real-camera calibration data. 5, 10, 15, and 20: outlier ratios in percent where the directory represents an outlier setting. Simulated sample names generally follow Sim_<image_size>_<region_id>. CAPD sample names retain the identifiers used during preprocessing. Data format All CSV files are headerless and comma-separated. Star and match identifiers are 1-based; 0 denotes an unmatched star. img.csv Each row describes a source detected in the image domain: Column Meaning Unit 1 Image-plane x coordinate pixel 2 Image-plane y coordinate pixel 3 Source brightness/intensity arbitrary unit 4 Image-source identifier 1-based integer 5 Matched catalogue-star identifier, or 0 1-based integer ast.csv Each row describes a reference catalogue star: Column Meaning Unit 1 Right ascension (RA) degree 2 Declination (Dec) degree 3 Apparent magnitude magnitude 4 Catalogue-star identifier 1-based integer 5 Matched image-source identifier, or 0 1-based integer The match columns in img.csv and ast.csv are reciprocal ground-truth correspondences. info.ini The [Info] section stores sample and camera metadata. Fields used by the released code include: Key Meaning image_size Image width/height in pixels. fov Field of view in degrees. f Camera focal length. dx Pixel pitch. RA, Dec Pointing center in degrees. ratio Outlier ratio used for the sample. level Distortion level. k1, k2, k3, p1, p2 Optional radial and tangential distortion coefficients. Distortion coordinate files Coordinate CSV files in distortion_data.tar.gz contain four columns: source_x, source_y, distorted_x, distorted_y Coordinates are expressed in pixels unless otherwise stated by the accompanying info.ini file. train, val, and test filenames indicate the split used by the distortion-correction experiments. Extraction ``` tar -xzf sim.tar.gztar -xzf match_subsets.tar.gztar -xzf capd.tar.gztar -xzf distortion_data.tar.gz``` Provenance and limitations The simulated data are semi-synthetic. Their noise, stray-light, dead-pixel, outlier, and distortion settings approximate representative wide-field observations but do not cover every camera or observing condition. capd.tar.gz contains processed coordinates, attributes, and correspondences derived from CAPD observations. It does not contain the original photographic-plate FITS images. Transferring a distortion-correction model to another camera requires camera-specific calibration or adaptation data. For the source CAPD project, please cite: Z.-J. Shang et al., "Digitization of Astronomical Photographic Plates of China and Astrometric Measurement of Single-exposure Plates," Research in Astronomy and Astrophysics, 24(5), 055010, 2024. https://doi.org/10.1088/1674-4527/ad339d