Didrien/meteorix-mendeley

Meteorix annotated meteor images

Bounding-box annotations of meteors on a subset of the Mendeley dataset Image dataset for the creation of an automatic system for meteor fall detection (Andre Gradvohl and Victor Yukio Shirasuna, 2023, DOI 10.17632/4k98n84d9g.1, CC BY 4.0). This is the dataset used to fine-tune and evaluate YOLO26n, RF-DETR Small and DETR-ResNet-50 in the Meteorix study.

Content

The original dataset provides 7,000 ground-based night-sky images (BRAMON, UKMON and BOAM networks) with image-level labels only. We kept all meteor images except one corrupted image (3,849), sampled 1,650 of the 3,150 meteor-free images at random, and drew one axis-aligned bounding box per meteor (single class meteor). The 5,499 images are split 70 / 10 / 20 %, with the same meteor / meteor-free ratio in every split:

Split	Images	With a meteor	Without
train	3,849	2,694	1,155
valid	550	385	165
test	1,100	770	330

The same images, splits and boxes are provided in two formats:

• yolo26.zip (folder yolo26/): YOLO format, images resized to 640x640 (<split>/images, <split>/labels, data.yaml), used for YOLO26n;
• coco.zip (folder coco/): COCO format, images resized to 800x800 (<split>/*.jpg, <split>/_annotations.coco.json), used for RF-DETR and DETR-ResNet-50.

manifest.sha256.json lists the SHA-256 checksum of every file inside the two archives.

from huggingface_hub import hf_hub_download
import zipfile

for name in ("yolo26.zip", "coco.zip"):
    zipfile.ZipFile(hf_hub_download("Didrien/meteorix-mendeley", name, repo_type="dataset")).extractall("data")

License and attribution

The images come from the dataset of Gradvohl and Shirasuna and are redistributed under its CC BY 4.0 license; our bounding-box annotations and splits are released under the same license. When using this dataset, please cite the original dataset:

Gradvohl, A., Shirasuna, V. Y. (2023). Image dataset for the creation of an automatic system
for meteor fall detection. Mendeley Data, V1. https://doi.org/10.17632/4k98n84d9g.1