Dataset Description and File Mapping

This is the official dataset for the manuscript "Detection of Strong Gravitational Lenses with Lens-Topology Relation and Arc-Assisted Annotations", accepted for publication in The Astrophysical Journal Supplement Series (ApJS). Sourced from DESI DR10, it contains 16,440 jitter-augmented images and corresponding object detection annotations. The dataset consists of 1 image archive and 3 annotation files, detailed as follows:

1. Image Data Directory: images/ (Extracted from images.zip, 13.43 GB)

Contains 16,440 Numpy array files (.npy format, loadable via numpy.load()). Each file is a tensor of shape (4, 256, 256), representing a 256x256 pixel image (0.262''/pixel) with 4 photometric bands (g, r, i, z). 

The file naming convention encodes specific metadata in the following format:
[Source_ID]__objra=[...]__objdec=[...]__cenra=[...]__cendec=[...]__[Jitter_ID]__ls-dr10__s=256.npy

- Sample Distinction (via [Source_ID]):
  Negative Samples (10,000 images): Prefixed with "NEG_". Includes normal galaxies and morphologically balanced hard negatives.
  Positive Samples (6,440 images): Not prefixed with "NEG_" (e.g., "A_", "candidate_"). The prefix indicates the original catalog source before cross-validation.

- Spatial Coordinates:
  "objra"/"objdec": True Right Ascension/Declination of the target in the source catalog.
  "cenra"/"cendec": RA/Dec of the absolute center of the cropped image.

- Jitter Augmentation (via [Jitter_ID]):
  Indicates which of the 5 position-jittered crops (e.g., "jitter0", "jitter1") this specific image represents for a given source.

2. Training Set Annotations: instances_train.json (4.50 MB)
COCO-format annotation file. Contains annotations for 11,505 images used for model training. Annotations include bounding boxes for both the "lens-system" (Category ID: 2) and the "lensed-arc" (Category ID: 1).

3. Test Set Annotations: instances_test.json (1.29 MB)
COCO-format annotation file. Contains annotations for 3,295 images, used for evaluating model generalization performance.

4. Validation Set Annotations: instances_val.json (659.77 kB)
COCO-format annotation file. Contains annotations for 1,640 images, used for validation and hyperparameter tuning during the training process.