Pith. sign in

REVIEW 3 major objections 7 minor 4 references

in-Car Biometrics (iCarB) Datasets for Driver Recognition: Face, Fingerprint, and Voice

T0 review · 3 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The iCarB project presents three new biometric datasets—face, fingerprint, and voice—collected inside a car from 200 consenting volunteers, claiming they are the largest and most diverse publicly available in-vehicle biometric datasets…

desk verdict Useful per-modality in-car biometric datasets, but the multimodal driver-recognition story is materially weakened because identities are deliberately unlinked across modalities. read the letter →

arxiv 2411.17305 v1 pith:CFSO6B6P submitted 2024-11-26 cs.CV

Vedrana Krivokuca Hahn , Jeremy Maceiras , Alain Komaty , Philip Abbet , Sebastien Marcel This is my paper · ORCID
classification cs.CV
keywords in-vehiclebiometricsdriverrecognitionfacefingerprintvoicemultimodalbiometricdatasetsdemographicbias
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper presents three biometric datasets collected inside a car from 200 consenting volunteers: face videos, fingerprint images, and voice samples. The authors claim these are the largest and most diverse publicly available in-vehicle biometric datasets, and that the fingerprint set is the first public in-car fingerprint dataset. The data was gathered in indoor and outdoor sessions with deliberately introduced variations such as masks, sunglasses, hats, dry or moist fingers, and background noise, to imitate real driver-recognition conditions. If these claims hold, the datasets give the research community a consent-based, multimodal benchmark for evaluating and comparing recognition systems and for studying demographic and environmental bias.

What carries the argument

The load-bearing mechanism is the data-collection protocol, which imposes controlled variations and records metadata per capture. Each subject is assigned an anonymous ID, and every file's name encodes the session, sensor, and variation, such as accessory and action for face, scanner and finger condition for fingerprint, and microphone, window state, and noise type for voice. The protocol also defines structured evaluation protocols in the form of enrolling and probing CSV files, so that recognition systems can be tested globally or per-variation. This design is what makes the three datasets usable as a benchmark and is the basis for the claimed breadth and diversity.

What would settle it

Compute verification accuracy on iCarB-Face's indoor, no-accessory, frontal protocol with any standard face recognition system; if the equal error rate is not clearly better than chance, the dataset would fail to support the claimed benchmarking utility.

Watch

Extended reading notes

Core claim

The central discovery is the datasets themselves: iCarB-Face, iCarB-Fingerprint, and iCarB-Voice, collected from 200 volunteers (100 male, 100 female) seated in the driver's seat, with skin tones spanning the full Fitzpatrick scale and ages from 18 to 60+. iCarB-Face contains 3,600 near-infrared face videos covering indoor and outdoor lighting, with and without masks, hats, and sunglasses, and with different head movements. iCarB-Fingerprint contains 9,521 fingerprint images from two scanners (one thermal, one optical) under normal, dry, moist, hot, and cold finger conditions. iCarB-Voice contains 36,000 voice samples recorded by two microphones under noiseless, traffic, music, discussion, and impulsive-noise conditions, with windows open and closed. The paper argues that because all three modalities share the same automotive environment and come with evaluation protocols and demographic and environmental metadata, the datasets enable unimodal and multimodal benchmarking, presentation-attack development, and bias analysis that existing single-modality in-vehicle datasets cannot.

Load-bearing premise

The claim that these datasets will be valuable for biometrics research assumes the collected data is actually usable for recognition benchmarking, yet the paper itself reports overexposed and underexposed face videos, missing fingerprint images for older and dry-skin subjects, and wind-distorted voice recordings, and it provides no baseline recognition experiments.

Editorial extensions

If this is right

  • Researchers can benchmark face, fingerprint, and voice recognition systems on data from the same automotive environment, using the provided per-variation protocols.
  • The datasets support construction of multimodal pseudo-identities for training and testing fusion algorithms, since IDs are not linked across modalities.
  • The included metadata enables demographic and environmental bias evaluations, for instance by skin type, age, gender, weather, or fingerprint condition.
  • The data can be used to generate presentation attacks for evaluating presentation-attack-detection algorithms on face, fingerprint, and voice modalities.
  • The fingerprint dataset, if confirmed as the first public in-vehicle one, fills a gap for automotive biometric research.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the datasets live up to their quality claims, they could serve as a common testbed for comparing unimodal versus multimodal driver recognition, which the field currently lacks.
  • Because the three datasets are not identity-linked across modalities, researchers cannot directly study cross-modal matching; the paper's implicit assumption that pseudo-identities suffice for fusion evaluation is a limitation worth testing.
  • A natural next step is to run baseline recognition experiments, which the paper does not provide, to establish the achievable performance on each protocol; without these, the claimed benchmarking value remains unquantified.
  • The deliberate noise conditions, such as dry fingers and windy voice captures, could be used to stress-test the robustness of recognition systems in ways that standard datasets do not.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript introduces three datasets (iCarB-Face, iCarB-Fingerprint, iCarB-Voice) of biometric data collected inside a car from 200 consenting volunteers per modality, with detailed capture protocols, metadata, and predefined evaluation protocols. The authors claim these are the largest and most diverse publicly available in-vehicle biometric datasets, with iCarB-Fingerprint being the first public in-vehicle fingerprint dataset. They also propose using the three datasets together via 'multimodal pseudo-identities' to train and test fusion algorithms. The paper provides file counts (3,600 face videos, 9,521 fingerprint images, 36,000 voice samples), internal consistency of protocols, and documented limitations.

Significance. If the datasets are usable for recognition, they would be a valuable resource for the biometrics community: they are ethically sourced (explicit consent, IRB approval), gender-balanced, span the Fitzpatrick scale, include multiple controlled and uncontrolled conditions, and are accompanied by protocols and metadata for bias and robustness studies. The explicit documentation of data quality limitations is commendable. However, the absence of any recognition baselines and the lack of cross-modal identity correspondence substantially weaken the central claims of benchmarking utility and multimodal driver recognition.

major comments (3)
  1. [Section I, Multimodality bullet; Section III.D] The identities of data subjects do not correspond across the three datasets, so the iCarB datasets do not support evaluation of genuine same-person multimodal (face+fingerprint+voice) recognition. The proposed 'multimodal pseudo-identities' pair unrelated subjects and are not validated as a substitute for real multimodal fusion; the paper gives no evidence that fusion behavior on pseudo-identities matches that on authentic multimodal data. This is a load-bearing weakness because the title and abstract frame the datasets as supporting driver recognition across three modalities. The authors should either release a subset with true cross-modal correspondences or substantially revise the multimodal utility claim.
  2. [Sections II and IV] No recognition experiments are reported for any of the three modalities. The paper asserts that the datasets can be used to evaluate and benchmark face, fingerprint, and voice recognition systems, but does not demonstrate that the data support such evaluation. Section IV documents overexposed and underexposed face videos, missing fingerprint images for older/dry-skin subjects, and wind-distorted voice recordings; without baseline results, a reader cannot judge whether standard matchers achieve usable accuracy on these data. At least one baseline experiment per modality (e.g., a face verification network, a minutiae-based fingerprint matcher, an i-vector/x-vector speaker verification system) is needed to substantiate the datasets' benchmarking value.
  3. [Abstract and Section I] The claim that iCarB constitute 'the largest and most diverse publicly available in-vehicle biometric datasets' is not substantiated. The comparison in Section I is limited to two datasets (VFPAD and 3DMAD) and only considers subject counts; no comprehensive survey of in-vehicle biometric datasets is provided, and 'diversity' is asserted from demographic metadata without quantitative measures. The claim should be supported by a systematic comparison with prior work, or softened to 'among the largest' with a precise statement of what was compared.
minor comments (7)
  1. [Section II.B] The word 'renumerated' should be 'remunerated'.
  2. [Section II.B (Voice)] The total of 180 voice samples per subject is initially confusing because the protocol lists 90 condition-sentence combinations; the text should explicitly state that each sentence was captured simultaneously by two microphones, yielding two files per sentence.
  3. [Section II.B and Section III] It is unclear whether the '200 volunteers' are the same individuals across the three datasets or three distinct cohorts. Section II says '200 volunteers for each piece of biometric data,' which suggests different subject pools; this should be clarified, as it affects the interpretation of the demographic statistics in Section III.
  4. [Section III.D] The phrase 'no intra-sensor variations were considered' is ambiguous; it likely means that each protocol uses data from a single sensor and does not mix sessions, but this should be restated more clearly.
  5. [Table III] For the microphones, the sampling rate and bit depth are not reported; these are essential for reproducible speaker recognition experiments.
  6. [Section II.A] The statement that the near-infrared camera has 'two illumination circuits' is not explained; specify whether they are used for different purposes (e.g., still frames vs. video) or are redundant.
  7. [Section I] The abbreviation 'VFPAD' is used without expansion; define it at first use (e.g., 'Vehicle Face Presentation Attack Detection').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the iCarB paper is a descriptive dataset contribution with no fitted predictions or self-citation chain that forces its central claims.

full rationale

This paper does not attempt a derivation, fitting, or prediction from first principles; it reports the construction and release of three biometric datasets. The central claims—largest and most diverse publicly available in-vehicle biometric datasets, first public in-vehicle fingerprint dataset, gender-balanced 200-subject cohort—are descriptive and externally checkable against the released data, not consequences of the paper's own assumptions. The subject-count arithmetic (3,600 face videos, 9,521 fingerprint images, 36,000 voice samples) follows from the described collection protocols and does not smuggle in a result as an input. The only self-citation, VFPAD [2], is used as a comparative example of an existing single-modality dataset, not as load-bearing authority for the paper's novelty claim. The paper's own Section I openly states that identities do not correspond across the three datasets, and Section III.D states that all provided protocols are per-modality and single-sensor; this is a genuine limitation for cross-modal driver-recognition benchmarking, but it is not circularity because the paper does not present genuine same-person multimodal recognition as a derived result. The limitations in Section IV (overexposed/underexposed face videos, missing fingerprint images, wind-distorted voice) further weaken the utility claim but again are not circular: they are disclosed quality issues, not fitted parameters renamed as predictions. No step reduces to its own input by definition or by self-citation, so a non-circular finding is appropriate.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claims rest on the assumption that the described collection and metadata are accurate and that the resulting data are usable for recognition research. No free parameters are fitted and no new entities are introduced.

assumptions (3)
  • domain assumption Volunteer recruitment through mailing lists and word of mouth produces a demographic distribution representative of the intended driver population.
    The paper claims high demographic diversity but the sample is a convenience sample, so representativeness is not guaranteed.
  • domain assumption The recorded metadata (gender, age category, Fitzpatrick skin type) are accurate.
    No independent verification or inter-rater reliability is reported.
  • domain assumption The simulated noise conditions approximate real in-vehicle biometric capture conditions.
    The data collection is designed to emulate non-ideal conditions; no comparison to actual driving capture is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of in-Car Biometrics (iCarB) Datasets for Driver Recognition: Face, Fingerprint, and Voice." pith.science (2026). https://pith.science/paper/CFSO6B6P

@misc{pith2026241117305,
  author       = {Pith},
  title        = {Pith review of: in-Car Biometrics (iCarB) Datasets for Driver Recognition: Face, Fingerprint, and Voice},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFSO6B6P}},
  note         = {Machine review of arXiv:2411.17305}
}
read the original abstract

We present three biometric datasets (iCarB-Face, iCarB-Fingerprint, iCarB-Voice) containing face videos, fingerprint images, and voice samples, collected inside a car from 200 consenting volunteers. The data was acquired using a near-infrared camera, two fingerprint scanners, and two microphones, while the volunteers were seated in the driver's seat of the car. The data collection took place while the car was parked both indoors and outdoors, and different "noises" were added to simulate non-ideal biometric data capture that may be encountered in real-life driver recognition. Although the datasets are specifically tailored to in-vehicle biometric recognition, their utility is not limited to the automotive environment. The iCarB datasets, which are available to the research community, can be used to: (i) evaluate and benchmark face, fingerprint, and voice recognition systems (we provide several evaluation protocols); (ii) create multimodal pseudo-identities, to train/test multimodal fusion algorithms; (iii) create Presentation Attacks from the biometric data, to evaluate Presentation Attack Detection algorithms; (iv) investigate demographic and environmental biases in biometric systems, using the provided metadata. To the best of our knowledge, ours are the largest and most diverse publicly available in-vehicle biometric datasets. Most other datasets contain only one biometric modality (usually face), while our datasets consist of three modalities, all acquired in the same automotive environment. Moreover, iCarB-Fingerprint seems to be the first publicly available in-vehicle fingerprint dataset. Finally, the iCarB datasets boast a rare level of demographic diversity among the 200 data subjects, including a 50/50 gender split, skin colours across the whole Fitzpatrick-scale spectrum, and a wide age range (18-60+). So, these datasets will be valuable for advancing biometrics research.

Figures

Figures reproduced from arXiv: 2411.17305 by the authors.

Figure 1
Figure 1. Biometric sensor set-up inside the car. The car cockpit was generated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 4
Figure 4. Microphones: (a) ME72; (b) AT2020. TABLE III MICROPHONE SPECIFICATIONS. Manufacturer Valeo Audio-Technica Model ME72 AT2020 Signal to Noise (dB(A)) >65 74 Current consumption (mA) 6 ± 3 2 [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Schematic of the circuit connecting the ME72 microphone to a battery [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: Fingerprint scanners: (a) OYSTER III; (b) CSD101i. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png]
Figure 7
Figure 7. Figure 7: Examples of face images in the Indoors scenario: (i) no accessories; [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]
Figure 8
Figure 8. Figure 8: Examples of face images in the Outdoors scenario: (i) no accessories; [PITH_FULL_IMAGE:figures/full_fig_p004_8.png]
Figure 9
Figure 9. Figure 9: Examples of face images where the data subject: (a) is still with [PITH_FULL_IMAGE:figures/full_fig_p004_9.png]
Figure 10
Figure 10. Figure 10: Examples of fingerprint images acquired with the Oyster III scanner. [PITH_FULL_IMAGE:figures/full_fig_p005_10.png]
Figure 11
Figure 11. Figure 11: Examples of fingerprint images acquired with the CSD101i scanner. [PITH_FULL_IMAGE:figures/full_fig_p005_11.png]
Figure 13
Figure 13. Figure 13: illustrates the directory structure for each dataset. Biometric data for each subject are stored in a dedicated folder labelled with the subject’s unique ID, referred to as <usr_id>, which ranges from 1 to 200. All data from both the indoor and outdoor sessions (see S…
Figure 12
Figure 12. Figure 12: Dataset statistics, with the skin colour distribution (based on the [PITH_FULL_IMAGE:figures/full_fig_p006_12.png]
Figure 14
Figure 14. Figure 14: Protocols folder structure. The protocol name indicates which variation is considered in the protocol; for example, in iCarB-Face, indoor-all will evaluate all the biometric samples that were acquired indoors, while indoor-mask-frontal will evaluate only the indoor sa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

  1. [1]

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

    work page

  2. [2]

    B. Lee, M. Hasegawa-Johnson, C. Goudeseune, S. Kamdar, S. Borys, M. Liu, and T. S. Huang, `` AVICAR: audio-visual speech corpus in a car environment ,'' in Interspeech, 2004, pp. 2489--2492

    work page 2004

  3. [3]

    Kotwal, S

    K. Kotwal, S. Bhattacharjee, P. Abbet, Z. Mostaani, H. Wei, X. Wenkang, Z. Yaxi, and S. Marcel, `` Domain-specific adaptation of CNN for detecting face presentation attacks in NIR ,'' IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 4, no. 1, pp. 135--147, 2022

    work page 2022

  4. [4]

    Jegham, A

    I. Jegham, A. B. Khalifa, I. Alouani, and M. A. Mahjoub, `` A novel public dataset for multimodal multiview and multispectral driver distraction analysis: 3MDAD ,'' Signal Processing: Image Communication, vol. 88, p. 115960, 2020

    work page 2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.