Read the paper
Discover the objects
→
Saliency-guided mining turns unlabeled scans into an object bank.
Localize Any Object in X-Ray Security Scanswithout Human Annotation
Explore the model
10 scans
Beyond familiar objects.
Real examples from the paper ↙Qualitative comparison
Cleaner boundaries and more complete instances, even under heavy overlap.
TOP-5 CONFIDENT MASKS
Original predictions from Figure 3. Grayscale scans make masks easier to inspect.
Performance across six benchmarks
Stronger box localization than SAM2, SAM3, and UnSAMv2 across all six test-only datasets.
SAM2SAM3UnSAMv2LAO-X (ours)
Method overview
Unlabeled scans become a world of structured training examples.
→Compose with physics
Ascene = Abg + ∑ Ai
Combine dual-energy scans in the Beer–Lambert absorbance domain.
Build up the challenge
2–4 objects0.05 max overlap
From sparse scenes to heavy overlap.
See the whole picture
Use a point and granularity to localize a part or a complete instance.
BibTeX
NeurIPS 2026@inproceedings{cai2026laox,
title = {Localize Any Object in {X-Ray} Security Scans without Human Annotation},
author = {Cai, Yaqi and Liu, Mingxuan and Vaquero, Lorenzo and Wang, Ning and Pu, Nan and Xue, Feng and Ricci, Elisa and Sebe, Nicu},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}