XSeg: A Large-scale X-ray Contraband Segmentation Benchmark For Real-World Security Screening

Hongxia Gao; Litao Li; Yixin Chen; Jiali Wen; Kaijie Zhang; Qianyun Liu

cs.CVApr 4, 2026

XSeg: A Large-scale X-ray Contraband Segmentation Benchmark For Real-World Security Screening

Authors: Hongxia Gao, Litao Li, Yixin Chen, Jiali Wen, Kaijie Zhang, Qianyun Liu

Organizations: Xi’an Jiaotong University · Shenzhen Loop Area Institute · Pazhou Laboratory · South China University of Technology

Abstract

X-ray contraband detection is critical for public safety. However, current methods primarily rely on bounding box annotations, which limit model generalization and performance due to the lack of pixel-level supervision and real-world data. To address these limitations, we introduce XSeg. To the best of our knowledge, XSeg is the largest X-ray contraband segmentation dataset to date, including 98,644 images and 295,932 instance masks, and contains the latest 30 common contraband categories. The images are sourced from public datasets and our synthesized data, filtered through a custom data cleaning pipeline to remove low-quality samples. To enable accurate and efficient annotation and reduce manual labeling effort, we propose Adaptive Point SAM (APSAM), a specialized mask annotation model built upon the Segment Anything Model (SAM). We address SAM's poor cross-domain generalization and limited capability in detecting stacked objects by introducing an Energy-Aware Encoder that enhances the initialization of the mask decoder, significantly improving sensitivity to overlapping items. Additionally, we design an Adaptive Point Generator that allows users to obtain precise mask labels with only a single coarse point prompt. Extensive experiments on XSeg demonstrate the superior performance of APSAM.

Figures & tables

Figure 1: XSeg data examples. The first row shows X-ray images of contraband from various sources, including but not limited to MobilePhone, Liquid, Scissors, Baton, and Handcuffs. Ground-Truth masks are generated by SAM [ 8 ] and refined by multiple security experts.

				Annotations
Dataset	Year	Classes	Positive	Classfication	Bbox	Mask	Caption	Type	Tasks
GDXray [ 14 ]	2015	3	8,150	✓	✓	✗	✗	Real	C, D
SIXray [ 15 ]	2019	6	8,929	✓	✓	✗	✗	Real	C, D
OPIXray [ 19 ]	2020	5	8,885	✓	✓	✗	✗	Synthetic	C, D
HiXray [ 17 ]	2021	8	45,364	✓	✓	✗	✗	Real	C, D
CLCXray [ 24 ]	2022	12	9,565	✓	✓	✗	✗	Real	C, D

Table 1: Comparison of X-ray Contraband Detection datasets, all datasets in the table are open source and available, Real means the data is from the real world, Synthetic means the data is from synthetic. ’C’,’D’,’S’,’ZS’ stand for Classification, Detection, Segmentation, and Zero-Shot tasks, respectively. XSeg has an advantage over the other datasets in terms of categories, the number of images that contain contraband.

Figure 2: Cross-domain chromatic variance phenomenon, the first row is PIDray [ 23 ] , the overall color is greenish, the second row is PIXray [ 13 ] , the imaging is still greenish, and the third row is HiXray [ 17 ] , the color distribution of the luggage is significantly different from PIDray and PIXray.

Figure 3: Pipeline of data cleaning. First, images are filtered based on resolution, aspect ratio, and noise levels using Laplacian variance thresholds, effectively removing low-resolution, irregularly proportioned, or excessively noisy samples. The cleaned images are then manually annotated by experts to establish high-precision ground truth masks. These annotations are subsequently used to train an in-house segmentation model, forming an automated labeling system. The model-generated masks undergo multiple rounds of iterative refinement to ensure boundary accuracy, with human verification at each stage. The final output is XSeg.

	Training	Validation	Test	Total
Xseg	68518	23735	10019	98644

Table 2: XSeg dataset split.

Figure 4: Framework of APSAM. APSAM initializes output tokens by concatenating maximum and minimum grayscale image representations. These are then processed by an Energy-Aware Encoder (EAE) and an MLP based Location Initializer. For the visual encoder backbone, we fine-tune with an identical adapter, keeping other parameters frozen. Point prompts are adaptively generated by our APG, and the mask decoder produces the final segmentation.

Figure 5: Energy-Aware Encoder and Location Initializer. The Energy-Aware Encoder comprises three convolutional layers featuring GELU activation and max-pooling. The Location Initializer primarily relies on linear layers and softmax for channel selection, ultimately employing Top-

k

filtering to identify the optimal token.

Methods	Backbone	mIoU(%)	Dice(%)	Trainable Params(M)	Pub’Year
PSPNet [ 25 ]	ResNet101	54.24	70.34	65.59	CVPR’17
DeepLabV3+ [ 1 ]	ResNet101	57.29	72.84	60.21	ECCV’18
UperNet [ 20 ]	Swin-Tiny	55.16	71.10	58.94	ECCV’18
Segformer [ 21 ]	MiT-b5	63.67	77.80	82.01	NIPS’21
Segmenter [ 16 ]	ViT-B/16	63.27	77.50	104.45	ICCV’21
Mask2former [ 2 ]	Swin-L	69.59	81.44	144.85	CVPR’22

Table 3: Comparison of APSAM with other segmentation methods, including Frozen variants, where all parameters are kept frozen for inference only, and Finetune counterparts, which are trained using the identical adapter strategy as APSAM.

Methods	mIoU(%)	Dice(%)
w/o EAE	71.90	81.62
w/o APG	70.89	79.50
w/o EAE & APG	67.87	77.45
w/ EAE & APG	72.83	82.31

Table 4: Ablation studies on our proposed EAE and APG demonstrate that optimal performance is achieved when both EAE and APG are utilized concurrently.

Prompt Type	mIoU(%)	Dice(%)
Random 1 point	67.87	77.45
Random 2 Point	70.31	80.18
APG (Ours)	71.90	81.62

Table 5: Ablation study of the APG, we evaluated the performance of randomly sampling a single point prompt, randomly sampling two point prompts, and our proposed APG. Our results clearly demonstrate that APG offers a significant advantage when utilizing two point prompts, outperforming simple random sampling.

Figure 6: Visualization of segmentation results. (g) shows the segmentation results of APSAM, which provides more complete and clearly visible segmentation boundaries than other schemes.

Methods	PIDray		PIXray
Methods	mIoU (%)	Dice (%)	mIoU (%)	Dice (%)
DeepLabV3+ [ 1 ]	64.62	73.34	73.17	86.36
Mask2former [ 2 ]	69.04	78.18	78.36	86.12
SegMAN [ 5 ]	66.51	75.24	78.37	87.87
SAM [ 8 ]	64.54	75.10	66.20	77.82
SAMUS [ 12 ]	67.01	76.85	80.63	88.26

Table 6: Evaluation results on PIDray and PIXray datasets. APSAM still maintains SOTA performance on the test sets of other datasets.

Figure 7: Display of the maximum and minimum grayscale images of contraband. The first column shows the contraband, the second column shows the corresponding maximum grayscale image, and the third column shows the corresponding minimum grayscale image.

Figure 8: Hierarchical structure of fine-grained semantic annotations in the Xseg dataset.

Superclass	Fine-grained Categories
Batteries	ColumnarBlockBattery, MotorBattery
Suspicious Liquid	PlasticbottleLiquid, GlassbottleLiquid, MetalbottleLiquid
Fruitknife	MetalhandleFruitknife, PlastichandleFruitknife
Cleaver	MetalhandleCleaver, PlastichandleCleaver
Scissors	MetalhandleScissors, PlastichandleScissors, Scissors
Batons	ExpandableBatons, Baton

Table 7: Fine-grained categories grouped by semantic superclasses.

Image Segmentation

Image segmentation assigns pixels or voxels to objects and regions, supporting measurement, scene understanding, and decisions across image domains.

See all 399 papers

OneFocus: Enabling Real-World X-ray Security Screening with a Unified Vision-Language Model
Jun 14, 2026Jiali Wen, Hongxia Gao, Litao Li +4Visual Question Answering Vision-Language Models
2South China University of Technology · 1Xi’an Jiaotong University · 3Shenzhen Loop Area Institute +1
Localize Any Object in X-Ray Security Scans without Human Annotation
Oct 5, 2026Yaqi Cai, Mingxuan Liu, Lorenzo Vaquero +5Self-Supervised Learning Synthetic Data Generation
University of Trento · Fondazione Bruno Kessler · Dalian University of Technology +1
FleXray: Universal Clinical X-ray Segmentation
Sep 22, 2026Victor Ion Butoi, Vivek Gopalakrishnan, John V. Guttag +2Medical Imaging Medical Image Segmentation
MIT CSAIL · Massachusetts General Hospital · Harvard Medical School

Jun 14, 2026J/K move · Enter open · S save

OneFocus: Enabling Real-World X-ray Security Screening with a Unified Vision-Language Model

Jiali Wen, Hongxia Gao, Litao Li, Yixin Chen+3

2South China University of Technology · 1Xi’an Jiaotong University · 3Shenzhen Loop Area Institute+1

Visual Question Answering Vision-Language Models Synthetic Data Generation

Open paper AskPDF arXiv

X-ray contraband detection is critical for security in large-scale logistics and transportation, yet conventional detectors struggle to adapt to emerging contraband types and lack fundamental visual understanding. Vision-language models (VLMs) offer strong generalization but are hindered by the scarcity of high-quality X-ray image-caption data. To bridge this critical gap, we present MMXray, a meticulously curated benchmark of 52,124 image-caption pairs spanning 28 fine-grained classes of X-ray contraband. To enrich MMXray with realistic occlusion patterns, we further introduce CleanDET, a dedicated synthesis dataset containing clean foreground contraband images from 28 categories and background images with diverse density levels, together with AnyContraSyn, a controllable synthesis method designed to operate on CleanDET. We also develop OnePipe, an extensible pipeline for systematic data curation. Built on MMXray, we propose OneFocus, a unified VLM that supports four core tasks: visual question answering, contraband localization, classification, and image understanding. OneFocus achieves state-of-the-art performance in X-ray contraband understanding and demonstrates robust cross-domain generalization, establishing a strong vision-language baseline for security screening.

Abstract

Image Segmentation

Explore similar work

OneFocus: Enabling Real-World X-ray Security Screening with a Unified Vision-Language Model

Localize Any Object in X-Ray Security Scans without Human Annotation

FleXray: Universal Clinical X-ray Segmentation