FrenchCastle/xray-baggages-customs

MIRO X-ray Security Scans: Baggage

109,944 baggage X-ray scans with 170,827 bounding boxes, gathered from 5 public X-ray security datasets and relabelled into one taxonomy, in COCO format, with per-image provenance and licence. The cargo scans live in the sibling repository FrenchCastle/miro-xray-cargo, with the same layout and taxonomy.

The collection was built for MIRO-VLM, a project that evaluates vision-language models as assistants to customs and security officers: reading a scan, finding threats and checking the goods against the declaration. Each source keeps its own split and its own terms; nothing is pooled across sources without a dataset field to tell them apart.

Licensing in one line. This is a redistribution of third-party data under their original terms, which differ per source; several are academic, non-commercial use only. Read Licensing before use and cite the original authors.

Quick start

Download everything (metadata and all image zips), then unpack:

hf download FrenchCastle/xray-baggages-customs --repo-type dataset --local-dir xray_baggage
cd xray_baggage && for z in images_*.zip; do unzip -q "$z" && rm "$z"; done

Or only one source:

hf download FrenchCastle/xray-baggages-customs --repo-type dataset --local-dir xray_baggage \
  --include "annotations/clcxray_coco.json" "images_clcxray.zip" MANIFEST.csv

Load it with pycocotools (any COCO reader works):

from pycocotools.coco import COCO

coco = COCO("xray_baggage/annotations/clcxray_coco.json")
img = coco.loadImgs(coco.getImgIds()[0])[0]  # file_name is relative to xray_baggage/
boxes = coco.loadAnns(coco.getAnnIds(imgIds=img["id"]))
print(img["file_name"], img["split"], [(coco.cats[b["category_id"]]["name"], b["bbox"]) for b in boxes])

The Hub dataset viewer is off: the images ship as zips so that each source can be fetched, cited and removed independently.

Dataset structure

Sources and splits

source	images	splits	boxes	images without boxes	zip
CLCXray	9,564	test 956, train 7,652, val 956	22,111	0	6.4 GB
DvXray	32,000	test 3,180, train 25,612, val 3,208	10,583	22,400	7.2 GB
HiXray	45,364	test 9,069, train 36,295	102,928	0	6.8 GB
OPIXray	8,885	test 1,776, train 7,109	8,919	0	2.6 GB
SIXray (Kaggle YOLO subset)	14,131	test 831, train 11,638, val 1,662	26,286	119	909.4 MB

Splits are the ones each source publishes (or, where a source has none, the one its distributor provides); they are not re-drawn. Use them per source. There is no cross-source test set, and no de-duplication across sources was attempted.

Files

path	content
annotations/<source>_coco.json	COCO detection file per source, unified class names
images_<source>.zip	images of that source, stored (uncompressed); members are images/<source>/..., exactly the COCO file_name
MANIFEST.csv	one row per image: image_id, file_name, dataset, domain, split, license, width, height, n_boxes
label_report.json	per source, original label to unified class
classes_mapping.json	the normalisation table and the unified class ids

COCO fields

Standard COCO detection, with extra keys:

• images[]: id, file_name, width, height, plus dataset (source key), domain (baggage), split (train / val / test) and license (the source's terms).
• annotations[]: id, image_id, category_id, bbox ([x, y, width, height], pixels), area, iscrowd, segmentation where the source provides masks or polygons, and attributes.original_label, the label as the source wrote it.
• categories[]: the full unified taxonomy, shared by both domain repositories, so class ids are stable across files. Classes absent from a file simply have no boxes.

Images without boxes are negatives (nothing of interest annotated), kept as the source ships them.

Taxonomy

Source labels are mapped to one taxonomy: threats and declarable items (knife_cutter, firearm, battery_powerbank, liquid_container...) and, for cargo, goods categories prefixed cargo_. The original label is always kept in attributes.original_label, so any other grouping can be rebuilt.

Class counts in this repository

class	boxes	images containing it
phone	53,835	37,430
battery_powerbank	20,935	18,017
knife_cutter	18,900	14,774
liquid_container	15,366	10,750
laptop	10,042	9,602
cosmetic	9,949	7,662
pliers	9,625	7,290
firearm	8,062	5,417
scissors	6,913	6,360
wrench	5,385	4,092
tablet	4,918	4,719
sprayer	1,776	1,650
lighter	1,650	1,553
hammer	749	741
baton	714	714
screwdriver	690	682
dart	660	652
fireworks	658	649

Label mapping

Original label to unified class, per source

source	original label	unified class
CLCXray	Cans	liquid_container
CLCXray	CartonDrinks	liquid_container
CLCXray	GlassBottle	liquid_container
CLCXray	PlasticBottle	liquid_container
CLCXray	SprayCans	sprayer
CLCXray	SwissArmyKnife	knife_cutter
CLCXray	Tin	liquid_container
CLCXray	VacuumCup	liquid_container
CLCXray	blade	knife_cutter
CLCXray	dagger	knife_cutter
CLCXray	knife	knife_cutter
CLCXray	scissors	scissors
DvXray	Bat	baton
DvXray	Battery	battery_powerbank
DvXray	Dart	dart
DvXray	Fireworks	fireworks
DvXray	Gun	firearm
DvXray	Hammer	hammer
DvXray	Knife	knife_cutter
DvXray	Lighter	lighter
DvXray	Pliers	pliers
DvXray	Pressure_vessel	sprayer
DvXray	Razor_blade	knife_cutter
DvXray	Saw_blade	knife_cutter
DvXray	Scissors	scissors
DvXray	Screwdriver	screwdriver
DvXray	Wrench	wrench
HiXray	Cosmetic	cosmetic
HiXray	Laptop	laptop
HiXray	Mobile_Phone	phone
HiXray	Nonmetallic_Lighter	lighter
HiXray	Portable_Charger_1	battery_powerbank
HiXray	Portable_Charger_2	battery_powerbank
HiXray	Tablet	tablet
HiXray	Water	liquid_container
OPIXray	Folding_Knife	knife_cutter
OPIXray	Multi-tool_Knife	knife_cutter
OPIXray	Scissor	scissors
OPIXray	Straight_Knife	knife_cutter
OPIXray	Utility_Knife	knife_cutter
SIXray (Kaggle YOLO subset)	Gun	firearm
SIXray (Kaggle YOLO subset)	Knife	knife_cutter
SIXray (Kaggle YOLO subset)	Pliers	pliers
SIXray (Kaggle YOLO subset)	Scissors	scissors
SIXray (Kaggle YOLO subset)	Wrench	wrench

Licensing

There is no single licence: each image keeps its source's terms, recorded in the license field of the COCO image and in MANIFEST.csv. Source datasets states the terms as each source publishes them. When in doubt, the upstream terms prevail, and anything beyond non-commercial research needs the source authors' permission. The annotation conversion and the label mapping added here are released under CC BY 4.0.

If you are an author of one of these datasets and want your data removed or its terms described differently, open a discussion on this repository and it will be handled promptly.

Source datasets

CLCXray (clcxray)

Cutters and liquid containers: subway-station scans plus scans of manually packed bags.

• Upstream: https://github.com/GreysonPhoenix/CLCXray
• Paper: https://ieeexplore.ieee.org/document/9722843
• Terms: Academic use only.
• Caveats: About half of the images are staged rather than stream-of-commerce.

DvXray (dvxray)

Dual-view baggage scans: 16,000 pairs (two orthogonal views), 15 prohibited item classes, positives and negatives.

• Upstream: https://github.com/Mbwslib/DvXray
• Paper: https://ui.adsabs.harvard.edu/abs/2024ITIF...19.3866M/abstract
• Terms: Academic use only.
• Caveats: Both views of a pair are separate images here; keep pairs on the same side of any split you make. Most images are negatives.

HiXray (hixray)

Real airport scans of passenger baggage with everyday items that must be declared (phones, laptops, power banks, cosmetics, liquids, lighters).

• Upstream: https://github.com/HiXray-author/HiXray
• Paper: https://openaccess.thecvf.com/content/ICCV2021/html/Tao_Towards_Real-World_X-Ray_Security_Inspection_A_High-Quality_Benchmark_and_Lateral_ICCV_2021_paper.html
• Terms: Academic use only; the authors ask for a signed commitment for formal use.
• Caveats: Heavily dominated by phones and power banks.

OPIXray (opixray)

Airport baggage scans with 5 kinds of cutters, many occluded.

• Upstream: https://github.com/OPIXray-author/OPIXray
• Paper: https://arxiv.org/abs/2004.08656
• Terms: Academic use only (OPIXray terms).
• Caveats: Only knife and scissors classes; every image contains an item.

SIXray (Kaggle YOLO subset) (sixray_yolo_subset)

A community YOLO-format subset of SIXray: subway baggage scans with guns, knives, wrenches, pliers and scissors.

• Upstream: https://www.kaggle.com/datasets/khanhbtq99/sixray
• Paper: https://openaccess.thecvf.com/content_CVPR_2019/html/Miao_SIXray_A_Large-Scale_Security_Inspection_X-Ray_Benchmark_for_Prohibited_Item_CVPR_2019_paper.html
• Terms: Academic use only (SIXray terms, https://github.com/MeioJane/SIXray).
• Caveats: Community re-annotation and re-split of SIXray, resized to at most 640 px; not the official SIXray release or split.

Dataset creation

Curation rationale. Public X-ray security datasets are scattered across Google Drive, Baidu, Kaggle, Roboflow and Hugging Face, in half a dozen annotation formats and label vocabularies. Evaluating a model across them first needs one format, one taxonomy and exact provenance for every image.

Processing. Each source was downloaded from its official or documented distribution, converted to COCO (from VOC XML, YOLO txt, paired txt or COCO), and its labels mapped to the unified taxonomy through a fixed table (classes_mapping.json). Images are the downloaded files, byte for byte: this repository does not resize, recompress or filter them (a distributor upstream may have, as the caveats note). Large sources may be a class-balanced subset, as stated in their caveats. Splits are kept as distributed.

Annotations. All boxes come from the source datasets (see each paper for its annotation protocol). This repository adds no new manual annotation; it renames labels and records the original.

Personal and sensitive information. The images are X-ray transmission scans of bags and cargo; they contain no faces, names or documents. Some sources are real stream-of-commerce scans from airports or subway stations; none is known to include personal data.

Considerations for use

Intended use. Research on detection and understanding of objects in X-ray security imagery: benchmarking detectors and vision-language models, studying domain shift between scanners and between baggage and cargo, and building tools that assist human inspectors.

Out of scope. Certifying or operating a screening system; any use that would help conceal items from X-ray inspection; commercial use of sources whose terms forbid it.

Known biases and limitations.

• Scanners, colour palettes and resolutions differ per source, and a model can learn the source instead of the object. Evaluate per source, or across sources on purpose.
• Class balance follows the sources: some classes come from a single source (and therefore a single scanner).
• Some sources are staged (items packed for the dataset) or synthetic (threats composited into real scans); these are not stream-of-commerce distributions.
• Near-duplicates exist in some sources (augmented copies, dual views); see caveats.
• Label granularity is reduced by the mapping (for example, every blade type becomes knife_cutter); use attributes.original_label for the finer label.

Citation

Cite the original datasets you use, and this consolidation if it helped:

@misc{chastel2026miroxray_baggage,
  title        = {MIRO X-ray Security Scans (Baggage): a unified COCO consolidation of public X-ray datasets},
  author       = {Chastel, Fran\c{c}ois},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/FrenchCastle/xray-baggages-customs}}
}

@article{zhao2022clcxray,
  title   = {Detecting Overlapped Objects in X-Ray Security Imagery by a Label-Aware Mechanism},
  author  = {Zhao, Cairong and Zhu, Liang and Dou, Shuguang and Deng, Weihong and Wang, Liang},
  journal = {IEEE Transactions on Information Forensics and Security},
  volume  = {17},
  pages   = {998--1009},
  year    = {2022}
}

@article{ma2024dvxray,
  title   = {Toward Dual-View X-Ray Baggage Inspection: A Large-Scale Benchmark and Adaptive Hierarchical Cross Refinement for Prohibited Item Discovery},
  author  = {Ma, Bowen and Jia, Tong and Li, Mingyuan and Wu, Songsheng and Wang, Hao and Chen, Dongyue},
  journal = {IEEE Transactions on Information Forensics and Security},
  volume  = {19},
  year    = {2024}
}

@inproceedings{tao2021hixray,
  title     = {Towards Real-World X-ray Security Inspection: A High-Quality Benchmark and Lateral Inhibition Module for Prohibited Items Detection},
  author    = {Tao, Renshuai and Wei, Yanlu and Jiang, Xiangjian and Li, Hainan and Qin, Haotong and Wang, Jiakai and Ma, Yuqing and Zhang, Libo and Liu, Xianglong},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {10923--10932},
  year      = {2021}
}

@inproceedings{wei2020opixray,
  title     = {Occluded Prohibited Items Detection: An X-ray Security Inspection Benchmark and De-occlusion Attention Module},
  author    = {Wei, Yanlu and Tao, Renshuai and Wu, Zhangjie and Ma, Yuqing and Zhang, Libo and Liu, Xianglong},
  booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},
  pages     = {138--146},
  year      = {2020}
}

@inproceedings{miao2019sixray,
  title     = {{SIXray}: A Large-scale Security Inspection X-ray Benchmark for Prohibited Item Discovery in Overlapping Images},
  author    = {Miao, Caijing and Xie, Lingxi and Wan, Fang and Su, Chi and Liu, Hongye and Jiao, Jianbin and Ye, Qixiang},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {2119--2128},
  year      = {2019}
}

Maintenance

Maintained by François Chastel (FrenchCastle). Report problems, missing credits or licence questions in the Community tab of this repository.