nnvff/Xray36-COCO-v1

Xray36 COCO v1

Snapshot of the completed 36-class X-ray object detection dataset, built on 2026-11-14 and uploaded on 2026-11-15. It combines OPIXray, HiXray, CLCXray, STCray, PIXray, PIDray, DvXray, and LDXray.

Split	Images	Bounding boxes
train	204,245	376,699
test	73,679	128,473
Total	277,924	505,172

Images retain their original bytes and resolution. All 277,924 images were checked against their recorded SHA-256 hashes during packaging. The source image payload is 43,746,932,972 bytes. The COCO annotation files are unchanged from the independently validated export.

Download and extract

This repository is public. Download it without authentication:

hf download nnvff/Xray36-COCO-v1 --repo-type dataset --local-dir Xray36_COCO_v1

Extract the image shards and provenance archives:

for archive in Xray36_COCO_v1/data/*.tar; do
  tar -xf "$archive" -C Xray36_COCO_v1
done
tar -xzf Xray36_COCO_v1/metadata.tar.gz -C Xray36_COCO_v1
tar -xzf Xray36_COCO_v1/reports.tar.gz -C Xray36_COCO_v1

Image shards are independent tar archives of roughly 2 GB each, separated by the existing train/test split. Each member retains its images/<source>/<original path> name. Extract all shards into the same directory. The archives are for faithful COCO download and extraction, not a WebDataset conversion.

annotations/instances_train.json and annotations/instances_test.json contain the original COCO annotations. Their file_name values are relative to the extracted images/ directory. Category IDs are 0–35 and bbox is [x, y, width, height]. Each annotation describes an individual object with iscrowd: 0.

To verify the downloaded repository files on Linux, run:

cd Xray36_COCO_v1
sha256sum -c SHA256SUMS

Verification can be run before or after extraction. manifest.json lists file sizes, SHA-256 hashes, per-shard image counts and split totals. class_statistics.csv gives the image and box counts for each category and split; images with multiple classes count once in each applicable category.

Contents

• data/train-*.tar, data/test-*.tar: every exported image, with original paths and bytes.
• annotations/: the two unchanged COCO annotation files.
• metadata.tar.gz: image and annotation manifests, copy verification, mapping, source scope, decisions, exclusions and other provenance.
• reports.tar.gz: process and visual reports, independent validation, COCO API checks and review evidence.
• scripts/: the original build and verification scripts, unchanged.
• README.original.md: the original build documentation, unchanged.

Key categories, mapping, summary and validation files are also available directly in metadata/ and reports/. Rebuild instructions in the original README refer to the original source workspace and require those source datasets. The upload excludes the environment, Python bytecode, scan cache, incomplete copy logs/status, execution logs and the redundant local report ZIP.

After extraction, open reports/process.html and reports/quality_report.html locally. Keep the extracted images/ directory alongside the reports so their image links work.

Dataset construction and limits

Original train/test membership is preserved, with CLCXray val included in train. DvXray contributes only OL top-view positive examples to train; LDXray contributes only A top-view images. PIDray uses original_2021. This version has no validation split and does not import records with no retained valid boxes.

Boxes were clipped to actual image bounds, zero-area boxes were removed, and 36 confirmed misaligned source images were excluded. STCray dimensions follow actual image dimensions without proportionally rescaling the original boxes. Original image pixels were not resized or cropped.

No deduplication was applied. Duplicates within and across splits remain, including 11 shared SHA-256 groups across train and test. Therefore this version does not establish strictly independent test images. Original annotation omissions, semantic errors and source-specific box conventions have not undergone exhaustive manual repair. The 10 passed independent checks verify compliance and export integrity; they are not a model performance evaluation.

Source rights

This is a research snapshot of eight third-party source datasets. Their respective terms and attribution requirements continue to apply. This upload does not assert a new unified license or public redistribution permission.