Abstract (short)
We study whether raw two‑dimensional gas chromatography / time‑of‑flight mass spectrometry (GC×GC–ToF‑MS) measurements can be used for identity verification without expert‑driven compound identification. Each measurement is converted into a compact multi‑channel image, and modern feature‑embedding methods are trained for verification. On a newly assembled human‑scent dataset of 252 individuals (2,528 samples, ~7.5 TB), the best method reaches ≈53% TPR at 5% FPR without explicit spatial registration. These results demonstrate feasibility for settings where cooperation or line‑of‑sight is limited, while leaving ample room for improvement.
Method at a glance
Channel representation
Local mass spectra are compressed into C overlapping, learnable channels (e.g., Gaussian kernels), yielding an image‑like tensor C × H × W suitable for CNNs.
Embedding & training
We evaluate classic descriptors, off‑the‑shelf CNN backbones (ResNet/VGG/Inception/GoogLeNet), and a lightweight custom CNN with anisotropic filters, trained using triplet loss.
Registration (optional)
Contrary to typical analytical pipelines, the best performance on our dataset is achieved without explicit canonical registration.
See the paper for details, ablations, and the evaluation protocol.
Resources
Citation
If you use this work, please cite:
@inproceedings{Spetlik2026ScentID,
title = {Identity Verification from Human Scent Using Channel Representation of 2D Gas Chromatography-Mass Spectrometry Data},
author = {Spetlik, Radim and Hlavsa, Jan and \v{C}echov\'a, Jana and Pojmanov\'a, Petra and Matas, Ji\v{r}\'{\i} and Urban, \v{S}t\v{e}p\'an},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2026}
}