DriveFace: A Cross-Spectral Through-Glass Face Dataset for On-the-Move Vehicular Border Control

Anjith George¹, Luis Luevano¹, Alain Komaty¹, Zeina Al Amine¹, Vidit Vidit¹, Sébastien Marcel^1,2

¹Idiap Research Institute, ²University of Lausanne (UNIL)

VIDAR PAX sensor-vehicle outdoor setup at 90 degrees

Sensor-vehicle outdoor acquisition setup: subjects are observed inside a vehicle through automotive glass using a near-infrared (NIR) sensor at a checkpoint, mirroring on-the-move vehicular border control.

Abstract

The continuous growth in cross-border mobility places increasing pressure on existing border control infrastructures, motivating on-the-move biometric authentication, in which travellers are identified directly inside their vehicles at checkpoints. Face recognition is well-suited to this setting, as it can be acquired passively and at a distance. Its development, however, is hindered by the lack of representative datasets: existing benchmarks are collected in controlled environments and do not capture the challenges inherent to vehicular acquisition, including motion blur, variable illumination, occlusions, and cross-spectral enrollment. To address this gap, we introduce DriveFace, a dataset for on-the-move face recognition in border-control scenarios, comprising NIR vehicle-crossing videos paired with smartphone-based pre-enrollment data. Baseline evaluations with state-of-the-art models show clear performance limitations under these realistic conditions, highlighting the need for dedicated methods to advance the field.

Highlights

70 consenting subjects captured over two sessions approximately two months apart, pairing visible-spectrum smartphone pre-enrollment with in-vehicle NIR probes.
Cross-spectral, through-glass acquisition: RGB references matched against NIR probes captured through automotive windows at varying tint and viewing angles, under stationary and moving conditions.
DriveFace-PAD presentation-attack subset with print, replay, and mask attacks under multiple tint and illumination levels.
Rich metadata (tint level, illumination, head pose, vehicle speed) and standardized FR and PAD protocols with strong baseline results.

The DriveFace Dataset

DriveFace captures the end-to-end biometric workflow at vehicular border crossings, enabling systematic study of cross-spectral matching, through-glass degradation, and adverse illumination within a single operationally representative benchmark. It comprises three components: (i) visible-spectrum pre-enrollment face captures from two smartphones (iPhone 12 and Samsung Galaxy S9), (ii) in-vehicle NIR face captures acquired through automotive glass, and (iii) presentation attacks.

Outdoor (45°)

Indoor car

Simulated tinted window

The three main probe capture settings in DriveFace.

Sample Data

Samples from a single subject: visible-spectrum reference captures from two sessions, real-world outdoor NIR probes acquired through clear vehicle windows at different viewing angles and motion conditions, and controlled indoor NIR probes through glass panels of increasing tint.

Reference (RGB) — two sessions

Session 1, hard profile

Session 2, frontal

Outdoor probes (NIR, clear window)

45°, moving

0°, stationary

Simulated tint indoor probes (NIR)

Clear window

Medium tint (T35)

Dark tint (T05)

Demographics

DriveFace is demographically diverse and gender-balanced (45.7% female, 54.3% male), spanning ages from 18 to 85 years. Apparent skin tone is summarized into light, medium, and dark groups (51, 11, and 6 subjects, respectively).

Age group distribution labeled by Fitzpatrick skin tones

Age group distribution labeled by Fitzpatrick skin-tone categories.

Face Recognition Results

We benchmark open-source face recognition models — AdaFace, LVFace, EdgeFace, and the cross-spectrally adapted xEdgeFace — across three operational protocols. Despite the RGB–NIR modality gap, robust models perform well outdoors, while the tinted-glass simulation protocol is the most challenging. Explicit cross-spectral adaptation (xEdgeFace) consistently improves over its EdgeFace backbone, with the largest gains in the hardest setting.

Protocol	Model	AUC	EER	VR@FAR=1%	Rank-1
Outdoor	AdaFace	99.45	2.69	96.01	97.42
	LVFace	98.50	5.26	90.17	92.26
	EdgeFace	98.38	5.37	91.35	93.29
	xEdgeFace	98.94	4.18	93.12	94.22
Simulation	AdaFace	96.23	8.26	88.27	90.46
	LVFace	95.19	11.42	79.68	83.95
	EdgeFace	94.09	11.73	82.93	84.72
	xEdgeFace	96.03	8.09	88.63	89.68
Indoor Car	AdaFace	98.59	3.11	96.24	97.11
	LVFace	98.52	5.29	90.23	93.27
	EdgeFace	98.75	4.48	93.92	93.34
	xEdgeFace	97.98	3.47	95.95	96.16

Face recognition performance (%) across the three evaluation protocols. Best VR@FAR=1% per protocol in bold.

Effect of Tint Level

On the controlled subset, recognition accuracy generally improves as visible light transmission increases, with clear glass giving the best results — directly quantifying the impact of window tint.

VLT	AdaFace	LVFace	EdgeFace	xEdgeFace
T05	98.36	98.72	98.42	98.20
T15	98.48	98.09	98.12	98.60
T20	99.31	99.04	98.23	98.72
T35	99.62	99.14	98.75	99.35
clear	100.00	99.86	99.87	100.00

AUC (%) across window tint levels.

Failure Cases

Most recognition failures occur under extreme tint, low contrast, extreme profile views, and occlusions.

Left (green): genuine pairs with the lowest match scores (missed true matches). Right (red): impostor pairs with the highest match scores (false acceptances). Top row: enrollment; bottom row: probe.

Presentation Attack Detection

The DriveFace-PAD subset covers print, replay, and mask attacks presented from inside the vehicle and captured through glass in NIR. Known attacks (grandtest) are detected reliably, but performance degrades sharply for unseen print and mask attacks, revealing a clear generalization gap.

Laser-matte print (T20)

Transparent mask (T05)

Silicone-mannequin (T20)

Laser-glossy, frontal

Paper mask, frontal

Resin mask, profile

Example presentation attacks from DriveFace-PAD under indoor and outdoor conditions.

Protocol	Model	APCER	BPCER	ACER
grandtest	DeepPixBiS	0.00	3.30	1.60
	CLIP (fc only)	4.40	2.60	3.50
	DinoV2 (fc only)	3.50	1.00	2.20
	EfficientNet-B0	0.10	0.90	0.50
	CLIP (full)	1.60	6.50	4.00
	ConvNeXtV2-Tiny	0.30	1.00	0.70
unseen_mask	DeepPixBiS	91.60	0.00	45.80
	CLIP (fc only)	78.80	0.00	39.40
	DinoV2 (fc only)	52.30	0.10	26.20
	EfficientNet-B0	54.50	0.00	27.20
	CLIP (full)	83.90	1.60	42.70
	ConvNeXtV2-Tiny	73.10	0.00	36.60
unseen_print	DeepPixBiS	36.70	0.00	18.40
	CLIP (fc only)	67.80	0.10	34.00
	DinoV2 (fc only)	53.00	0.20	26.60
	EfficientNet-B0	70.70	0.00	35.40
	CLIP (full)	64.70	3.40	34.00
	ConvNeXtV2-Tiny	42.80	0.10	21.50

PAD error rates (%) at the EER threshold. Best ACER per protocol in bold.

BibTeX

@inproceedings{george2026driveface,
  title={DriveFace: A Cross-Spectral Through-Glass Face Dataset for On-the-Move Vehicular Border Control},
  author={George, Anjith and Luevano, Luis and Komaty, Alain and Al Amine, Zeina and Vidit, Vidit and Marcel, S\'ebastien},
  booktitle={IEEE International Joint Conference on Biometrics (IJCB)},
  year={2026},
}