CoChat
Sign Up
Cited by 73Face recognition and analysisBiometric Identification and SecurityFace and Expression Recognition

The DOI matches its Crossref record; 27 of the 27 references it cites were found in a registry or an index; none of the 18 sources its claims cite is checked for support yet.

Longitudinal Study of Child Face Recognition

Open access
Deb, Debayan, Nain, Neeta, Jain, Anil K.·2018·2018 International Conference on Biometrics (ICB)
DOI 10.1109/icb2018.2018.00042
Open this paper in CoChat — free accountOpen PDF

Abstract

We present a longitudinal study of face recognition performance on Children Longitudinal Face (CLF) dataset containing 3,682 face images of 919 subjects, in the age group [2,18] years. Each subject has at least four face images acquired over a time span of up to six years. Face comparison scores are obtained from (i) a state-of-the-art COTS matcher (COTS-A), (ii) an open-source matcher (FaceNet), and (iii) a simple sum fusion of scores obtained from COTSA and FaceNet matchers. To improve the performance of the open-source FaceNet matcher for child face recognition, we were able to fine-tune it on an independent training set of 3,294 face images of 1,119 children in the age group [3,18] years. Multilevel statistical models are fit to genuine comparison scores from the CLF dataset to determine the decrease in face recognition accuracy over time. Additionally, we analyze both the verification and open-set identification accuracies in order to evaluate state-of-the-art face recognition technology for tracing and identifying children lost at a young age as victims of child trafficking or abduction.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

This paper presents a longitudinal study of child face recognition on the Children Longitudinal Face (CLF) dataset, containing 3,682 face images of 919 subjects aged 2–18 years. It evaluates COTS-A, fine-tuned FaceNet, and their score fusion, finding that recognition accuracy degrades with elapsed time but that identification of missing children remains feasible over several years.

Key findings

  • Fused matcher achieved the best verification performance: 90.18% TAR at 0.1% FAR after 1 year, degrading to 73.33% after 3 years; COTS-A achieved 81.94% and 49.33%, and FaceNet 83.77% and 59.80%, at the same FARs and time lapses.
  • In open-set identification at 1% FAR, Fused achieved 77.86% Rank-1 and 79.01% Rank-3 after 1 year, and 75.39% Rank-1 and 76.42% Rank-3 after 7 years; DIR remained stable beyond rank 3.
  • Genuine scores decreased by 0.2234, 0.2180, and 0.2444 standard deviations per year for COTS-A, FaceNet, and Fused, respectively, corresponding to one full standard deviation after 4.5, 4.6, and 4.1 years.
  • The study estimates that 80% of the CLF population can be successfully recognized at 0.1% FAR for 2.5, 2, and 2.5 years by COTS-A, FaceNet, and Fused, respectively; 99% remain above threshold at 0.01% FAR for 2.5, 2, and 2.5 years.
  • Gender effects were matcher-independent: average genuine scores were not statistically different between boys and girls, but slopes were significantly steeper for boys, and girls appeared easier to recognize with higher overall genuine scores.

Methods

  • The CLF dataset contains 3,682 face images of 919 children aged 2–18 years, with an average of 4 images per subject collected over an average time span of 4.2 years; demographic makeup is 604 boys (66%) and 315 girls (34%).
  • Face comparison scores were obtained from COTS-A, FaceNet, and a simple sum fusion of their scores; FaceNet was fine-tuned on an independent Child Face Training (CFT) dataset of 3,294 face images of 1,119 children aged 3–18 years.
  • Longitudinal verification used 2,763 genuine comparison scores between enrollment and subsequent images, with 3.38 million impostor scores; open-set identification used 919 enrollment images in the gallery and 3,520 probe images, including 756 subjects not in CLF.
  • Multilevel statistical models with time-lapse and gender covariates were fit to standardized genuine scores using the LME4 package in R with maximum likelihood estimation; non-parametric bootstrapping with 1,000 bootstrap sets was used to obtain confidence intervals.

Limitations

  • The paper states that the CLF dataset cannot be released publicly due to privacy issues, though FaceNet similarity scores can be obtained from the authors for repeatability studies.
  • The paper states that the CLF dataset violates the normality assumption of multilevel models, so non-parametric bootstrapping was performed; it also describes the study as “only a small step” and calls for larger child face datasets and periodic longitudinal studies.
  • Evident from the text: only one COTS matcher and one open-source matcher are evaluated, and the lack of public dataset release limits independent replication.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/icb2018.2018.00042

Cites

27 of 27 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 of 18 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The CLF dataset contains 3,682 face images of 919 subjects in the age group [2,18] years.

    “We present a longitudinal study of face recognition performance on Children Longitudinal Face (CLF) dataset containing 3, 682 face images of 919 subjects, in the age group [2, 18] years.”
    • Also stated in: § 2. Children Longitudinal Face (CLF) dataset · plain
      “The Children Longitudinal Face (CLF) dataset contains 3, 682 face images of 919 children, in the age range of 2 to 18 years.”
    • Also stated in: § 4. Conclusions · plain
      “We obtained the Children Longitudinal Face (CLF) dataset containing 3,682 face images of 919 children in the age group of [2,18] years with an average of 4 images per subject collected over an average time span of 4.2 years.”
    Quote in the paperSubstantial claim
  • Findingnone given

    Each subject in CLF has at least four face images acquired over a time span of up to six years.

    “Each subject has at least four face images acquired over a time span of up to six years.”
    Quote in the paperSubstantial claim
  • Findingown data

    CLF images were collected over an average time span of 4.2 years.

    “We obtained the Children Longitudinal Face (CLF) dataset containing 3,682 face images of 919 children in the age group of [2,18] years with an average of 4 images per subject collected over an average time span of 4.2 years.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Verification accuracy for Fused decreased from 90.18% to 73.33% for a time lapse of 3 years.

    “Table 3 found that the verification accuracy for Fused decreased from 90.18% to 73.33% for a time lapse of 3 years.”
    • Also stated in: § 4. Conclusions · plain
      “Face verification accuracy for a time lapse of 1 year is high (TAR of 90.18% at 0.1% FAR for Fused), but degrades to 73.33% TARs at 0.1% FAR after 3 years of elapsed time between enrollment and probe image of a child.”
    Quote in the paperSubstantial claim
  • Findingown data

    Fused face matcher has the best verification performance over time: 90.18% TAR at 0.1% FAR with 1 year time lapse, compared to COTS-A 81.94% and FaceNet 83.77%.

    “Fused face matcher has the best verification performance over time (90.18%TAR at 0.1% FAR with 1 year time lapse), compared to COTS-A (81.94% TAR at 0.1% FAR with 1 year time lapse) and FaceNet (83.77%TAR at 0.1% FAR with 1 year time lapse).”
    Quote in the paperSubstantial claim
  • Probably not a substantial claim. 13
  • It may state two separate claims as one. 9

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    CLF face images were captured at resolution 354 × 472 pixels.

    “The face images were captured with a resolution of 354 × 472 pixels (Figure 3).”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Each subject in CLF has an average of 4 images acquired over an average time lapse of 4 years (minimum 2 years; maximum 7 years).

    “Each subject has an average of 4 images acquired over an average time lapse of 4 years (minimum time lapse of 2 years; maximum time lapse of 7 years).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    CLF demographic makeup is 604 (66%) boys and 315 (34%) girls.

    “Demographic makeup of CLF dataset is comprised of 604 (66%) boys and 315 (34%) girls.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingnone given

    In the verification scenario, the impostor distribution includes all possible impostor comparisons, totaling 3.38 million scores.

    “In the verification scenario, the impostor distribution includes all possible impostor comparisons, totaling 3.38 million scores.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Identification performance decreases with an increase in time lapse, but the rate of degradation in identification accuracy over time is very low.

    “Similar to the verification scenario, identification performance decreases with an increase in time lapse, however, the rate of degradation in identification accuracy over time is very low.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Genuine scores for COTS-A, FaceNet, and Fused decrease by one full standard deviation after 4.5, 4.6, and 4.1 years of time lapse.

    “Therefore, genuine scores for COTS-A, FaceNet, and Fused decrease by one full standard deviation of their respective score distribution after 4.5, 4.6, and 4.1 years of time lapse.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim.

    Findingnone given

    For open-set identification, 756 subjects with one face image per subject (not in CLF) were included in the probe set, totaling 3,520 probe images.

    “Additionally, we included 756 subjects with one face image per subject (not in the CLF dataset) in our probe set for open-set identification, totaling 3,520 probe images.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingreasoning

    The CLF dataset is time-unstructured and unbalanced because time lapse and number of image acquisitions per subject vary.

    “Time lapse between a probe and enrollment image and number of image acquisitions per subject in the CLF dataset varies from subject to subject and therefore, the dataset is time-unstructured and unbalanced.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingown data

    In CLF, genuine scores total 5,946 and impostor scores total 3.38 million.

    “Genuine scores (total of 5,946 scores) are computed as all pairwise comparisons between face images of the same subject and impostor scores are comprised of all possible impostor comparisons (total of 3.38 million scores) in the CLF dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    For longitudinal evaluation, enrollment images compared to subsequent images yield 2,763 genuine comparison scores.

    “face images for each subject at enrollment (first image acquisition) are compared to subsequent face image acquisitions of the same subject (a total of 2,763 genuine comparison scores).”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Premisecitation

    CACD does not include subjects younger than 10 years old, and only 199 subjects are present below age 18.

    “The Cross-Age Celebrity Dataset (CACD) [28] was collected to evaluate face recognition performance under aging, but subjects younger than 10 years old are not included in this dataset, and only 199 subjects are present below the age of 18.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim.

    Premisecitation

    Best-Rowden et al. found state-of-the-art face recognition technology has a very low TAR of 47.93% at 0.1% FAR for the age group [0,4] years.

    “Their results show that state-of-the-art face recognition technology has a very low True Accept Rate (TAR) of 47.93% at 0.1% False Accept Rate (FAR) for this age group of [0, 4] years.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim.

    Premisecitation

    Face recognition accuracy degrades with an increase in time lapse between gallery and probe image acquisitions.

    “Limited studies related to aging have indeed shown that (i) accuracy of face recognition degrades with an increase in time lapse between a subject's gallery and probe image acquisitions [23], [4], [24]”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [23]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [4]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [24]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Face recognition accuracies for older subjects are higher than younger ones.

    “and (ii) face recognition accuracies for older subjects are higher than younger ones [24], [25].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [24]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [25]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Premisereasoning

    PCSO, LEO, and MSP datasets are comprised of subjects above age 18 and are not suitable for a study focusing on tracing missing children.

    “However, these datasets are comprised of subjects above the age of 18 and are not suitable for our study which focuses on tracing missing children.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    For all three face matchers, girls appear to be easier to recognize than boys with higher genuine scores overall.

    “Therefore, for all three face matchers, girls appear to be easier to recognize than boys with higher genuine scores overall.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim.

    Study design and conductown data

    Non-parametric bootstrapping was conducted with 1,000 bootstrap sets by sampling all 919 subjects with replacement.

    “By sampling all the 919 subjects in the dataset with replacement, non-parametric bootstrapping is conducted with 1,000 bootstrap sets.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Premisecitation

    Ricanek et al. achieved verification accuracy of 24% at 0.1% FAR and closed-set identification performance of 25% on ITWCC.

    “A verification accuracy of 24% at 0.1% FAR was achieved, whereas closed-set identification performance was only 25%.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Premisereasoning

    Face recognition is perhaps the primary biometric modality for tracing missing children because parents and relatives are more likely to have a lost child's photographs than fingerprint or iris.

    “To trace missing children, face recognition is perhaps the primary biometric modality since parents and relatives are more likely to have a lost child's photographs(s) as opposed to, say, fingerprint or iris.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Deb et al. report genuine scores of 99.0% of the population remain above threshold at FAR 0.01% for elapsed time 10.5 years for COTS on PCSO and MSP.

    “Deb et al. [7] report that genuine scores of 99.0% of the population remain above the threshold at a FAR of 0.01% for an elapsed time of 10.5 years for a state-of-the-art COTS face matcher on both the PCSO and MSP datasets.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim.

    Study design and conductown data

    Non-parametric bootstrapping was performed to obtain confidence intervals for parameter estimates.

    “CLF dataset violates this parametric assumption of normality and therefore, non-parametric bootstrapping is performed to obtain confidence intervals for the parameter estimates [10].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [10]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • The authors state this as an assumption: Multilevel statistical models assume that residual errors are normally distributed.
21 of 21

  • The authors state this as an assumption.

    Multilevel statistical models assume that residual errors are normally distributed.

    “Multilevel statistical models are based on the assumption that the residual errors are normally distributed.”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Deb et al. found that the accuracy of face recognition decreases over time, with 83.77% TAR at 0.1% FAR with a 1-year time-lapse, decreasing to 59.80% TAR after 3 years.

    “In this study, the researchers found out that the accuracy of face recognition decreases over time.Initially, they achieved an accuracy of 83.77% True Acceptance Rate (TAR) and 0.1% False Acceptance Rate (FAR) with a 1-year time-lapse, but this decreased to 59.80% TAR after 3 years.”

    From Longitudinal Evaluation of Child Face Recognition and the Impact of Underlying Age (2024) · cites this paper as [6]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • CLF minimum image count conflicts with the abstract

    Major

    The abstract says every subject has at least four images, but Table 2 gives a range beginning at two images per subject for CLF.

  • Verification time lapses misidentify Table 3 columns

    Minor

    The text says Table 3 reports verification at three and five years, but its verification columns are for one and three years.

  • NITL image total exceeds the stated per-subject maximum

    Minor

    Table 2 lists 3,144 images for 314 NITL subjects, averaging about 10 images per subject-more than its stated maximum of five.

    Reported 5 · recomputed 10.012739

  • Conflicting maximum time spans for CLF

    Minor

    The abstract gives a maximum span of six years, whereas the dataset description gives a maximum of seven years.

  • One acquisition session conflicts with annual acquisitions

    Minor

    The dataset criteria say each subject has only one image acquisition session, while the CLF figure caption describes images of subjects acquired annually.

5 of 5

Citations

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. 0, Facenet Github source code
  2. 0, A life revealed
  3. 0, How the afghan girl was identified by her iris patterns
  4. brierley, 2015, A Long Way Home A Memoir
  5. davis, 2016, LION
  6. flanagan, 2017, Face recognition prize challenge, NIST
  7. 2017, The scandal of the missing children abducted from India's railway stations
  8. 0, Convention on the Rights of the Child., UN General Assembly

Cited by

Sorted by citations

Related

Metadata

Authors
Deb, D., Nain, N., & Jain, A. K.
Affiliations
  • Michigan State University (US)
  • Malaviya National Institute of Technology Jaipur (IN)
ORCID
Venue
2018 International Conference on Biometrics (ICB)
Publisher
IEEE
Published
2018-02
Pages
225-232
DOI
10.1109/icb2018.2018.00042
OpenAlex
W2768443437
Cited by
73
References
26
Citations by year
2026: 4 · 2025: 10 · 2024: 8 · 2023: 9 · 2022: 12 · 2021: 10 · 2020: 12 · 2019: 7 · 2018: 1
FWCI
2.20
Type
conference paper
Language
en
Open access
Licence
public-domain
Locations
  • International Conference on Biometrics (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, public-domain)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
  • Physical Sciences › Computer Science › Signal Processing › Biometric Identification and Security
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face and Expression Recognition
Keywords
score-level fusion, open-set recognition, verification accuracy, child trafficking, child abduction

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review