CLCXray: X-Ray Illicit Object Detection Benchmark

• CLCXray is a public benchmark for illicit object detection in X-ray imagery, defined by 12 classes, frequent object overlaps, and a mix of real and simulated subway scans.
• The dataset comprises 9,565 images with detailed bounding-box annotations, featuring a split that emphasizes an internal train-test shift and includes liquid-container and cutter categories.
• Comparative studies using mAP metrics reveal that transformer-based detectors outperform custom CNN architectures in handling large, overlapped objects despite higher computational costs.

CLCXray is a public benchmark for illicit object detection in X-ray security imagery. In comparative evaluation literature, it is described as an object detection dataset with 12 classes, 9,565 images, bounding-box annotations, RGB pseudo-colored imagery, and dual-energy X-ray acquisition, and it is used to study detection under frequent overlap among objects, including liquid-container categories that are stated to be absent from other datasets (Cani et al., 23 Jul 2025). A separate study on edge-oriented threat detection likewise treats CLCXray as an open-source X-ray security baggage dataset from the real subway security baggage screening process, converted into YOLO format for 12-class object detection (Ambati et al., 2023).

1. Definition and benchmark role

CLCXray is identified in the benchmarking literature as the dataset introduced by Zhao et al. for “Detecting Overlapped Objects in X-Ray Security Imagery by a Label-Aware Mechanism.” It is positioned as a benchmark designed to address two limitations of earlier X-ray security datasets: insufficient overlap among foreground objects and neglect of liquid containers. In the six-dataset comparison comprising OPIXray, CLCXray, SIXray, EDS, HiXray, and PIDray, CLCXray is treated as the benchmark most directly associated with object overlap/occlusion with same-class instances and background clutter, rather than with scanner-domain shift or tiny deliberately hidden objects (Cani et al., 23 Jul 2025).

This role is technically consequential. The comparative study characterizes CLCXray as a regime in which detectors must cope with overlapped, often large, foreground items, including both cutters and liquid containers. Relative to OPIXray, it is described as less about explicit occlusion-level stratification and more about frequent overlaps among multiple foreground objects; relative to PIDray, it occupies the opposite size regime, because PIDray contains only small objects whereas CLCXray is overwhelmingly large-object; relative to EDS, it does not target domain shift across scanners (Cani et al., 23 Jul 2025).

2. Dataset composition, classes, and split structure

CLCXray combines real subway security inspection imagery with simulated baggage setups. Of the 9,565 total images, 4,543 are real samples collected from subway scan inspection systems, and 5,022 are simulated by scanning artificially designed baggage arrangements. The dataset is split into training (80%), validation (10%), and test (10%) subsets; one implementation-oriented study gives the corresponding counts as 7,652 training images, 955 validation images, and 955 test images (Cani et al., 23 Jul 2025, Ambati et al., 2023).

Category	Classes
Cutter	Blade (BL), Dagger (DA), Knife (KN), Scissors (SC), Swiss army knife (SW)
Liquid-container	Can (CA), Carton drink (CD), Glass bottle (GB), Plastic bottle (PL), Vacuum cup (VA), Spray can (SP), Tin (TI)

A dataset-specific and methodologically important detail is that the test set uses a 1:9 real-to-simulated ratio, whereas the training and validation sets use an 8:1 real-to-simulated ratio (Cani et al., 23 Jul 2025). This suggests an internal train–test composition shift rather than a purely i.i.d. split. One paper additionally states that the data were pre-processed and converted into YOLO format, but does not provide a more detailed annotation schema, class remapping policy, or relabeling procedure (Ambati et al., 2023).

3. Overlap structure, object scale, and class difficulty

The dataset is explicitly described as overlap-heavy. On average, each image contains more than two potentially dangerous items, and nearly 60% of images contain at least two or more foreground objects. In the evaluated test partition used in the comparative study, CLCXray contains 1,421 total test objects, of which 3 are small, 225 medium, and 1,193 large (Cani et al., 23 Jul 2025).

These statistics shape the interpretation of detector behavior. The comparative analysis argues that CLCXray favors architectures that can recognize large objects under overlap-rich conditions. The same study reports that Knife (KN) is among the easier object types, whereas Glass bottle (GB) is among the harder ones. The broader explanation given is based on material attenuation and signature complexity: high-density metallic objects generally produce stronger contrast in X-ray imaging, while low-density or less distinctive liquid-container objects remain difficult (Cani et al., 23 Jul 2025).

Within the material-complexity discussion, Carton drink (CD) and Glass bottle (GB) are listed among low-attenuation low-complexity signatures, while Vacuum cup (VA) and Spray can (SP) fall into higher-complexity or pressurized-container categories (Cani et al., 23 Jul 2025). This suggests that CLCXray is not merely an overlap benchmark; it is also a benchmark in which class difficulty varies with X-ray attenuation properties and container morphology.

4. Evaluation protocol and comparative detector behavior

The unified comparative study evaluates CLCXray with mAP$^ {50}$ and mAP$^ {50:95}$ for detection quality, and inference time (ms), parameter size (M), and GFLOPS for efficiency and computational complexity. The definitions are given as follows:

$$\text{IoU} = \frac{|A \cap B|}{|A \cup B|} \in [0,1],$$

$\text{AP}=\int_0^1 p(r)\,dr \in [0,1],$

$\text{mAP}=\frac{1}{N}\sum_n^N \text{AP}_n,$

where $N$ is the number of classes. The study further specifies that mAP$^ {50}$ uses IoU threshold 0.5, while mAP$^ {50:95}$ averages AP across IoU thresholds from 0.5 to 0.95 in steps of 0.05 (Cani et al., 23 Jul 2025).

For CLCXray, YOLOv8 and RT-DETR are trained using SGD, with learning rate 0.01 and momentum 0.9; the custom CNN variants CHR, DOAM, and LIM use the same optimizer and hyperparameters. DINO uses AdamW, learning rate 0.0001, $\beta_1=0.9$, and weight decay 0.0001; Co-DETR uses AdamW, learning rate 0.0002, $\beta_1=0.9$, and weight decay 0.0001. YOLOv8, RT-DETR, CHR, DOAM, and LIM are trained for 100 epochs, while DINO and Co-DETR are trained for 36 epochs; early stopping is used throughout (Cani et al., 23 Jul 2025).

Detector	mAP$^ {50:95}$0	mAP$^ {50:95}$1
D(Co-DETR, Swin-B)	0.772	0.654
D(YOLOv8, Next-ViT-S)	0.740	0.640
D(YOLOv8, CSPDarkNet53)	0.733	0.636

Across all ten detectors, the best overall detector under both metrics is D(Co-DETR, Swin-B), with mAP$^ {50:95}$2 and mAP$^ {50:95}$3. The comparative study interprets this as evidence that transformer and hybrid detectors with transformer backbones outperform CNNs on CLCXray, “mainly due to the efficiency of transformer blocks in recognizing large objects.” The same analysis also reports a negative result: the custom X-ray-specific CNN modules CHR, DOAM, and LIM all underperform the generic YOLOv8 baseline on CLCXray, despite the dataset’s original emphasis on overlap handling (Cani et al., 23 Jul 2025).

The efficiency trade-off is explicit. The most accurate model, D(Co-DETR, Swin-B), is also the slowest and heaviest among the leading contenders, at 187 ms inference time, 125M parameters, and 1068 GFLOPS. By contrast, D(YOLOv8, CSPDarkNet53) reaches 0.636 mAP$^ {50:95}$4 with 7.52 ms, 43.6M parameters, and 165.4 GFLOPS, while D(YOLOv8, Next-ViT-S) reaches 0.640 mAP$^ {50:95}$5 with 12.22 ms, 56M parameters, and 174.9 GFLOPS (Cani et al., 23 Jul 2025).

5. Resource-constrained detection and QCFS Tiny YOLOv7

A distinct line of work uses CLCXray as the target benchmark for resource-constrained X-ray baggage threat object detection. That study evaluates three related detector families: a conventional Tiny YOLOv7 ANN, a QCFS Tiny YOLOv7, and a partially converted SNN Tiny YOLO (Ambati et al., 2023).

The architectural change in the QCFS model is narrowly defined. The authors modify Tiny YOLOv7 by replacing the standard activation in BConv with QCFS and replacing MaxPool with AvgPool. The rationale is ANN-to-SNN conversion theory: the paper states that when the shift hyperparameter $^ {50:95}$6, the expected conversion error is zero,

$^ {50:95}$7

The best QCFS result reported on CLCXray is obtained at 4 quantization steps (Ambati et al., 2023).

After 50 epochs, the paper reports approximate validation-style results of [email protected] ≈ 0.813, [email protected]:.95 ≈ 0.628, and F1 = 0.80 for Tiny YOLOv7, versus [email protected] = 0.836, [email protected]:.95 ≈ 0.649, and F1 = 0.82 for QCFS Tiny YOLOv7. On the held-out test set, the reported values are 0.616 [email protected] and 0.60 F1 for Tiny YOLOv7, versus 0.633 [email protected] and 0.62 F1 for QCFS Tiny YOLOv7 (Ambati et al., 2023).

The same study also evaluates partial spiking conversion. When the first QCFS activation is converted to a neuron, performance drops to [email protected] < 0.01; when the last QCFS activation is converted, it reaches 0.375 [email protected] (Ambati et al., 2023). This indicates that, within that study, the practical deployment winner is the QCFS ANN, not the converted spiking model. The paper claims that QCFS Tiny YOLO achieves state-of-the-art performance on CLCXray, but it does not provide a formal comparison table against prior published CLCXray methods, so the strongest directly demonstrated conclusion is narrower: QCFS outperforms the authors’ own standard Tiny YOLOv7 baseline on CLCXray (Ambati et al., 2023).

6. Reproducibility, limitations, and interpretive issues

CLCXray is comparatively well specified at the benchmark-summary level and less well specified at the low-level data-pipeline level. In the large comparative study, the framework code and model weights are made publicly available, and the implementation is said to rely on the authors’ released code together with Ultralytics and MMDetection; backbone initialization is reported as COCO-pretrained for CSPDarkNet53 and HGNetV2, ImageNet-22k-pretrained for Swin-B, and ImageNet-1k-pretrained for Next-ViT-S (Cani et al., 23 Jul 2025). In the Tiny YOLO study, the transfer-learning starting point is MS COCO pretrained Tiny YOLOv7 weights (Ambati et al., 2023).

At the same time, several reproducibility gaps are explicit. The comparative study does not provide CLCXray-specific resizing, normalization constants, augmentation recipes, label cleaning, annotation correction, relabeling, batch size, input resolution, scheduler, warmup policy, or mixed-precision details; the Tiny YOLO study likewise omits image resolution statistics, resizing policy, normalization specifics, augmentation details, anchor recalculation, class balancing strategy, and hardware-level deployment measurements such as latency, FPS, memory usage, power, or energy (Cani et al., 23 Jul 2025, Ambati et al., 2023).

A recurrent interpretive issue is leaderboard comparability. The benchmark-scale study evaluates ten detectors under a common protocol and emphasizes that detector choice in X-ray screening should be dataset- and challenge-aware rather than architecture-dogmatic (Cani et al., 23 Jul 2025). By contrast, the resource-constrained study reports strong QCFS results but does not present a formal cross-paper comparison table, so its “state-of-the-art” claim is not fully substantiated within the paper itself (Ambati et al., 2023). This suggests that CLCXray should be treated less as a single settled ranking and more as a benchmark whose conclusions depend on protocol, metric, and deployment constraint.

In the arXiv literature considered here, CLCXray is therefore best understood as a specialized benchmark for overlap-heavy, mostly large-object X-ray threat detection. Its distinctive properties are the coexistence of same-class/background overlap, multiple foreground objects per image, predominantly large instances, and liquid-container categories. Those properties explain why attention-based models can perform strongly on it, why low-attenuation liquid objects remain challenging, and why it remains useful as a complement to other public X-ray benchmarks rather than a substitute for them (Cani et al., 23 Jul 2025).