OPIXray: Occluded X-Ray Benchmark

• OPIXray is a specialized X-ray object detection benchmark focusing on occluded prohibited items, particularly five cutter-related classes.
• The dataset comprises 8,885 pseudo-colour dual-energy X-ray images with synthetic insertion of real luggage backgrounds and professional annotations.
• It serves as a testbed for evaluating detection architectures, attention modules, and training strategies under varying occlusion levels.

Searching arXiv for recent and foundational papers on OPIXray and related benchmarks/methods. OPIXray, short for Occluded Prohibited Items X-ray, is a baggage X-ray object detection benchmark centered on occluded prohibited items in security inspection imagery. Introduced by Wei et al. as the first high-quality object detection dataset for this setting, it has been used to study cutter-like threat detection under overlap, clutter, synthetic occlusion, detector architecture choice, training strategy, lossy compression, and cross-benchmark evaluation. Across the literature, OPIXray is consistently defined by its focus on five cutter-related categories, its official train/test split, and its explicit subdivision of the test set by occlusion severity (Wei et al., 2020, Webb et al., 2021, Cani et al., 23 Jul 2025).

1. Dataset identity and composition

OPIXray contains 8,885 X-ray images and focuses on five prohibited item categories: Folding Knife (FO), Straight Knife (ST), Scissor (SC), Utility Knife (UT), and Multi-tool Knife (MU). The original benchmark reports 7,109 training images and 1,776 testing images, with all images stored as JPEG at resolution 1225 × 954. The category totals reported in the benchmark are 1,993 folding knives, 1,044 straight knives, 1,863 scissors, 1,978 utility knives, and 2,042 multi-tool knives (Wei et al., 2020).

Later work describes OPIXray as RGB, dual-energy, and pseudo-colour X-ray imagery, and emphasizes that it is a detection benchmark for widely occurring prohibited items in baggage screening. One comparative evaluation further characterizes the test set as containing 1,772 objects, all “medium”, for COCO-style size analysis (Webb et al., 2021, Cani et al., 23 Jul 2025).

Property	Value
Total images	8,885
Modality	RGB / pseudo-colour dual-energy X-ray
Resolution	1225 × 954 JPEG
Classes	FO, ST, SC, UT, MU
Split	7,109 train / 1,776 test
Annotation	Bounding boxes
Test stratification	OL1, OL2, OL3

The benchmark is often summarized as approximately one instance per image, and one later survey states that it has 8,885 instances over 8,885 images. The original benchmark is more precise: 35 images contain more than one prohibited item, with 30 such images in training and 5 in testing (Cani et al., 23 Jul 2025, Wei et al., 2020). This suggests that OPIXray is predominantly, but not absolutely, a single-instance benchmark.

2. Construction, annotation, and imaging assumptions

The original OPIXray paper states that the dataset uses real luggage backgrounds scanned by airport X-ray machines, while the prohibited items are synthesized into those backgrounds using professional software used in airport training systems. Annotations were performed by professional security inspectors from an international airport, and the annotation workflow was double-checked using standards described as similar to PASCAL VOC quality control (Wei et al., 2020).

Subsequent papers describe OPIXray as an “Artificial Synthesis” dataset in which prohibited objects are synthetically placed into luggage images to create occlusion, while also stressing that the scenes are meant to mimic the randomly stacked and overlapping reality of personal baggage. Another comparative study frames the collection context as an international airport, with security inspectors asked to simulate real-world cases in which objects are stacked in luggage to produce varying degrees of object occlusion (Tao et al., 2021, Cani et al., 23 Jul 2025).

The benchmark therefore occupies a distinctive position in the X-ray literature. It is neither a purely naturalistic stream of unmodified checkpoint imagery nor a purely abstract synthetic dataset. A plausible implication is that OPIXray was designed to preserve operational baggage appearance while giving researchers explicit control over occlusion difficulty.

3. Evaluation protocol and occlusion stratification

A defining feature of OPIXray is the subdivision of its test set into three occlusion regimes:

• OL1: no or slight occlusion
• OL2: partial occlusion
• OL3: severe or full occlusion

The original benchmark reports the following test-set counts by occlusion level: 922 instances in OL1, 548 in OL2, and 306 in OL3. Per category, OL1 contains 206 FO, 88 ST, 160 SC, 214 UT, 255 MU; OL2 contains 148 FO, 84 ST, 126 SC, 88 UT, 105 MU; OL3 contains 50 FO, 63 ST, 83 SC, 41 UT, 70 MU (Wei et al., 2020). Later work retains the same three-way structure, sometimes writing the subsets as OP1, OP2, and OP3 (Hassan et al., 2021).

Most OPIXray papers report mAP at IoU = 0.5, following VOC-style detection evaluation. One operational study makes the AP definition explicit as

$\text{AP} = \int_0^1 p(r)\,dr,$

with mAP defined as the mean AP over the five OPIXray classes (Webb et al., 2021). More recent unified evaluations also report mAP$^{50}$ and mAP$^{50:95}$, and provide results separately for OL1, OL2, OL3, and the overall test set (Cani et al., 23 Jul 2025).

Because the literature mixes VOC [email protected], mAP$^{50}$, mAP$^{50:95}$, different image sizes, and different epoch schedules, direct numerical comparison across papers requires care. This is not a contradiction in the benchmark itself; it reflects evolving evaluation practice around the same dataset.

4. OPIXray as a methodological testbed

OPIXray has been used to evaluate several distinct classes of methods. In the original benchmark paper, Wei et al. proposed the De-occlusion Attention Module (DOAM), a plug-and-play module combining Edge Guidance and Material Awareness to refine detector features for occluded X-ray objects (Wei et al., 2020). A later extension added an over-sampling training strategy and reported results under the name DOAM-O, explicitly targeting hard, high-occlusion samples (Tao et al., 2021).

Tao et al. used OPIXray to evaluate the Lateral Inhibition Module (LIM), a feature-pyramid plug-in composed of Bidirectional Propagation (BP) and Boundary Activation (BA). In that work, OPIXray served as the complementary benchmark to HiXray for testing detectors under synthetic occlusion scenarios, and LIM was compared directly against DOAM on OPIXray (Tao et al., 2021).

Operational detector studies have treated OPIXray as a benchmark for architecture and deployment trade-offs. One such study compared Cascade R-CNN and FreeAnchor, examined JPEG compression robustness, and systematically evaluated augmentations such as RandomCrop, MixUp, BboxMixUp, CutMix, and ClassCutMix on OPIXray (Webb et al., 2021).

OPIXray has also become a testbed for DETR-like anti-overlap designs. AO-DETR introduced Category-Specific One-to-One Assignment (CSA) and Look Forward Densely (LFD) on top of DINO, explicitly targeting foreground-background coupling and edge blurring in X-ray imagery (Li et al., 2024). MMCL added Multi-Class Min-Margin Contrastive Learning to deformable DETR-based detectors by regularizing decoder content queries (Li et al., 2024). CSPCL further aligned content queries with classifier-weight prototypes using Intra-Class Truncated Attraction (ITA) and Inter-Class Adaptive Repulsion (IAR) losses (Li et al., 28 Jan 2025).

Beyond supervised detection, OPIXray has supported alternative formulations. One unsupervised study treated prohibited items as anomalies relative to normal baggage content and evaluated a Gaussian-Weighted Fourier Stylization (GW-FS) plus encoder-decoder anomaly segmentation pipeline on OPIXray without retraining the reconstruction network on OPIXray itself (Hassan et al., 2021). Another line of work used OPIXray for tensor-pooling contour instance segmentation, emphasizing multi-scale contours rather than anchor or query design (Hassan et al., 2021).

5. Empirical findings on OPIXray

The original benchmark established strong early baselines. On OPIXray, SSD achieved 70.89 mAP, YOLOv3 achieved 78.21, and FCOS achieved 82.02. With DOAM, these rose to 74.01, 79.25, and 82.41, respectively. Across occlusion levels, SSD+DOAM reached 77.87 on OL1, 72.45 on OL2, and 70.78 on OL3, compared with 75.45, 69.54, and 66.30 for plain SSD; the relative gain increased with occlusion severity (Wei et al., 2020). DOAM-O later reported 74.57 for SSD, 80.40 for YOLOv3, and 83.80 for FCOS, with particularly large gains on the Straight Knife category (Tao et al., 2021).

Later architectural studies pushed the reported OPIXray numbers higher. In the operational comparison of Cascade R-CNN and FreeAnchor, FreeAnchor with a ResNet-50 backbone achieved 87.7 mAP on OPIXray, compared with 85.9 for Cascade R-CNN and 82.4 for FCOS+DOAM (Wei et al.) in the comparison table. The same study reported that RandomCrop gave the largest single augmentation gain on OPIXray, improving FreeAnchor by +3.65 mAP and Cascade R-CNN by +3.64, while MixUp gave +0.9 for FreeAnchor and +0.8 for Cascade R-CNN. Under JPEG compression, FreeAnchor dropped only from 87.7 to 87.2 at quality 50 when trained on uncompressed images, and retraining on compressed images raised performance at quality 10 from 77.4 back to 87.4 (Webb et al., 2021).

DETR-based models produced another round of gains. Under a unified 320 × 320, 12 epoch setup against general detectors, AO-DETR reached 79.2 mAP with ResNet-50 and 80.8 with Swin-L, exceeding the corresponding DINO baselines of 78.2 and 80.0. In a higher-capacity OPIXray comparison against specialized prohibited-item detectors, AO-DETR reached 87.2 with ResNet-50 and 89.0 with Swin-L at 640 × 640 and 15 epochs (Li et al., 2024).

Contour-driven methods also remained competitive. The tensor-pooling contour instance segmentation framework reported 0.8396 mAP on OPIXray, with occlusion-level scores of 0.7946 on OP1, 0.7382 on OP2, and 0.7291 on OP3, surpassing both TST and FCOS+DOAM in that comparison (Hassan et al., 2021). In the unsupervised anomaly-detection formulation, the reconstruction-based system reported 0.7483 mAP on OPIXray, slightly below TST at 0.7532 but above DOAM at 0.7401 and DOAM-O at 0.7457; it also reported F1 = 0.6560 against 0.3074 for GANomaly and 0.3464 for Skip-GANomaly (Hassan et al., 2021).

The broadest later comparison came from a six-dataset unified evaluation. On OPIXray, D(DINO, Swin-B) and D(Co-DETR, Swin-B) achieved the highest overall mAP$^{50}$, both at 0.928, while D(YOLOv8, Next-ViT-S) achieved the highest mAP$^{50:95}$ at 0.429. For the occlusion subsets, D(YOLOv8, Next-ViT-S) had the highest mAP$^{50:95}$ on OL1, OL2, and OL3, with 0.444, 0.424, and 0.417, respectively. The same study also reported that D(YOLOv8+DOAM, CSPDarkNet53) and D(YOLOv8+LIM, CSPDarkNet53) underperformed the plain D(YOLOv8, CSPDarkNet53) baseline on OPIXray, despite earlier gains reported for DOAM and LIM in older detector families (Cani et al., 23 Jul 2025).

6. Position in the X-ray literature

OPIXray occupies a specialized but influential niche in X-ray security research. Compared with HiXray, it is narrower in category diversity but more explicitly occlusion-focused; compared with SIXray, it is a true object detection benchmark rather than an image-level classification resource; compared with EDS and PIDray, it is not designed to probe domain shift or hidden-item behavior as directly. One 2025 evaluation therefore uses OPIXray as the benchmark “for occlusion,” alongside other datasets used for domain shift, hidden objects, or broader category diversity (Tao et al., 2021, Cani et al., 23 Jul 2025).

A common misconception is to treat OPIXray as a general-purpose proxy for all X-ray prohibited-item detection. The benchmark is explicitly limited to five cutter-related classes, and multiple papers note its relatively small scale and restricted threat diversity. One operational study states that OPIXray has a “relatively small number of samples and limited diversity of threat categories,” while later comparisons emphasize that it is knife-dominated and does not include the broader threat space covered by datasets such as HiXray or PIDray (Webb et al., 2021, Cani et al., 23 Jul 2025).

Another recurrent issue is methodological transfer. Earlier papers showed clear gains from DOAM, DOAM-O, and LIM when added to detectors such as SSD, YOLOv3, FCOS, SSD, FCOS, and YOLOv5 (Wei et al., 2020, Tao et al., 2021, Tao et al., 2021). The later unified evaluation, however, found that legacy X-ray-specific modules could degrade performance when transplanted into YOLOv8, and explicitly described this as a potentially “counter-intuitive” outcome (Cani et al., 23 Jul 2025). This suggests that OPIXray is not only a benchmark for occlusion robustness but also a benchmark for architectural compatibility.

The benchmark’s continuing relevance lies in that role. It remains a standard site for testing whether a detector, feature pyramid, attention mechanism, query-assignment strategy, contrastive regularizer, or anomaly-localization system can handle overlap, partial visibility, and boundary ambiguity in pseudo-colour baggage X-ray imagery. At the same time, the literature consistently indicates that OPIXray is best interpreted as one component of a broader evaluation suite rather than as a complete stand-in for the operational X-ray screening problem.