X-ray security scans skip human labels: mining objects from unlabeled scans lifts cluttered-bag localization by up to 23 points
Synopsis
Universal object localization in X-ray security screening has been limited by scarce expert annotation, forcing reliance on RGB-pretrained segmentation models that falter on overlapping, weakly bounded baggage scans. LAO-X mines object instances from unlabeled X-ray scans, recomposes them in the absorbance domain under the Beer-Lambert law to build cluttered scenes with holistic and part pseudo-masks, then fine-tunes a SAM2 localizer through an occlusion-controlled curriculum. Across six X-ray benchmarks, it raises category-agnostic mAP over SAM2 and X-ray-specific baselines by 2% to 23% with no human-annotated labels, reaching 53.33% mAP on DET-COMPASS.
Interpretation
Object localization in X-ray screening can be done without any human-annotated masks, boxes, points, or category labels, with the model learning objectness rather than category names or text prompts. Prior X-ray detectors rely on expert-annotated predefined categories, and open-vocabulary methods still need category names or visual exemplars at inference, while RGB foundation models transfer poorly to X-ray imagery. On six test-only X-ray benchmarks, LAO-X reaches 63.39% AP50, 55.73% AP75, and 53.33% mAP on DET-COMPASS, outperforming SAM2, SAM3, UnSAMv2, and RAXO baselines.
Training supervision comes from the unlabeled scans themselves: object instances are mined, recomposed into cluttered baggage scenes under X-ray transmission physics, and paired with both holistic instance and local part pseudo-masks. Earlier unsupervised segmentation relies on RGB-style visible-surface grouping, which struggles with projective superposition, blurred boundaries, and low-attenuation targets in X-ray scans. On PIXray and PIDray, LAO-X achieves the best mIoU and F1, exceeding the strongest baseline by more than 10 points on PIXray and reaching 64.06% and 71.86%.
An occlusion-controlled curriculum lets an RGB-pretrained segmentation model adapt progressively to X-ray superposition, moving from sparse objects to severe overlap. Training once on mixed-complexity synthetic scenes is less effective, yielding 16.49% mAP compared with the staged curriculum. Raising object counts and overlap caps through Easy, Hard, and Extreme stages lifts mAP from 15.58% to 16.83% and finally to 18.33%.
Perspective
The method is validated on six test-only X-ray benchmarks for category-agnostic localization, with all training supervision drawn from unlabeled COMPASS-XP scans and physics-guided synthetic scenes, using no target-domain annotations. The intended readers are researchers and engineers working on X-ray security perception, foundation-model domain adaptation, and unsupervised segmentation. The source positions LAO-X as an assistive localization front end for downstream recognition and human inspection, and practical use requires validation and calibration on the target scanner with human oversight retained.
Low-attenuation objects remain a difficult regime: on the invisible subset of DET-COMPASS, AP50 is 26.69%, well below the 63.39% on the full set, since thin plastic, fabric, or paper boundaries may be hard to separate from background. On the PIDray Hidden subset, LAO-X has the highest AP and AP75 but an AR10 of 14.60, below SAM3 at 31.35 and UnSAMv2 at 19.59, so stronger top-ranked coverage does not mean all instances are recovered when more predictions are allowed. The synthesis does not explicitly simulate scattering, beam hardening, detector noise, calibration differences, or scanner-specific pseudo-color mapping, so robustness across scanners and operating conditions is not guaranteed.