Pith. sign in

REVIEW 4 major objections 4 minor 10 references

Surveying Facial Recognition Models for Diverse Indian Demographics: A Comparative Analysis on LFW and Custom Dataset

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A new Indian student face dataset shifts how recognition models rank, lifting LDA+kNN from 75.66% to 87.76%.

desk verdict A small new Indian face dataset is the only real contribution; the benchmark numbers are not reliable as reported because the split may leak identities and the authors call the results 'hypothetical'. read the letter →

arxiv 2412.08048 v1 pith:BIKNZJNC submitted 2024-12-11 cs.CV cs.LG

Pranav Pant , Niharika Dadu , Harsh V. Singh , Anshul Thakur This is my paper · ORCID
classification cs.CVcs.LG
keywords facialrecognitionIndiandemographicsJFADLFWCNNhybridmodeldatasetbias
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

The paper introduces a small custom dataset of Indian student faces, called JFAD, and compares a range of face-recognition models on it against the standard LFW dataset. The authors report that the CNN and hybrid models reach near-perfect accuracy on both datasets, and that most models, especially LDA+kNN, score higher on JFAD than on LFW (LDA+kNN rises from 75.66% to 87.76%). They conclude that dataset selection materially changes recognition performance and that building inclusive, population-specific datasets is important for fair deployment. The work also offers the JFAD collection as a resource for future evaluation.

What carries the argument

The load-bearing objects are the two datasets and the comparison protocol. JFAD, the new custom set, contains 400 images of 40 Indian students (ten per subject) captured under varied real-world conditions; the LFW subset mirrors that structure with 50 subjects and ten images each. After face detection, augmentation (rotation, shear, brightness/contrast/sharpness), and resizing, the images are split 80:20 into training and test sets, and each model is evaluated on both datasets. The model suite spans subspace projections (PCA, LDA, LPP), Gabor-filter features, a convolutional network, and a hybrid that combines Haar and edge features with a convolutional classifier. Running the same models on the two datasets is the mechanism that yields the paper's conclusions about dataset selection.

What would settle it

Re-run the same models with a subject-disjoint 80:20 split (all ten images of each subject on one side) and check whether CNN accuracy on JFAD stays near 98.4% and LDA+kNN stays near 87.76%; if accuracy drops substantially, the original numbers were inflated by identity leakage.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the choice of evaluation dataset changes both the absolute accuracy and the ranking of face-recognition models. Using a newly collected set of 400 Indian student portraits (JFAD, expanded to 2,816 images after augmentation) alongside an LFW subset (3,508 images), the authors find that the CNN and a hybrid model that combines Haar and edge features with a convolutional network achieve near-perfect accuracy on both. The linear-discriminant model (LDA+kNN) improves markedly on JFAD, from 75.66% to 87.76%, and the hybrid model attains perfect accuracy on both datasets. The authors interpret these results as evidence that dataset composition—ethnic diversity, capture conditions, and image resolution—strongly affects recognition performance, and that population-specific benchmarks are needed to tune and validate systems for deployments in India.

Load-bearing premise

The reported accuracies assume the 80:20 train/test split keeps every person's images entirely in either training or testing, but the paper never states that the split is by identity.

Editorial extensions

If this is right

  • If the results hold, model rankings on LFW cannot be assumed to transfer to Indian populations; the same model can appear weak or strong depending on the benchmark.
  • The LDA+kNN jump from 75.66% to 87.76% suggests that classical linear methods can become competitive when the training and test faces come from a homogeneous, population-specific set.
  • The CNN and hybrid models' near-perfect accuracy on both datasets implies that current deep architectures handle this Indian student sample at least as well as the LFW sample.
  • The authors' conclusion that dataset selection is critical supports building and sharing more region-specific benchmarks and tuning systems per deployment context.

Reading between the lines

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

  • As an editorial extension, because the paper does not state that its 80:20 split keeps all images of a subject on one side, the near-perfect accuracy figures may partly reflect within-person similarity rather than generalization to unseen identities.
  • With 40 subjects in JFAD and 50 in LFW, the reported accuracy differences carry large sampling uncertainty, so a broader Indian cohort would be needed to confirm the ranking shifts.
  • The dataset is drawn from a narrow academic age band (18–23), so treating JFAD as representing India's full ethnic and demographic diversity would overstate the paper's evidence.
  • A direct test would be to re-evaluate all models on a subject-disjoint split and on a more varied multi-region Indian dataset to see whether the LDA+kNN jump and near-perfect CNN/hybrid scores persist.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper evaluates several face recognition approaches—PCA+kNN, LDA+kNN, LPP, Gabor+kNN, a CNN, and a hybrid Haar/edge+GoogLeNet model—on two datasets: a 50-subject subset of LFW and a newly collected 40-subject Indian student dataset (JFAD). It reports that the CNN and hybrid models achieve near-perfect accuracy on both datasets, and that most models perform better on JFAD than on the LFW subset, with LDA+kNN rising from 75.66% to 87.76%. The authors conclude that dataset selection significantly affects face recognition performance and that the JFAD is a useful benchmark for Indian demographics.

Significance. If the empirical claims were properly supported, the paper would offer a useful contribution: a new Indian-focused face dataset and a comparison of classical versus deep models on it. The motivation to study demographic diversity in face recognition is timely and important. However, the evaluation as described does not establish the central quantitative results. The train/test split is not stated to be identity-disjoint, the results are reported as single runs without error bars, and Section 6 refers to the reported numbers as 'hypothetical results.' These issues affect every comparative claim in Section 5 and prevent the paper from making a trustworthy empirical contribution in its current form.

major comments (4)
  1. [Section 4.4 and Section 5.3] The experimental procedure does not state that the 80:20 split is identity-disjoint. Section 4.4 says only 'The datasets were then split into training and testing sets with a 80:20 ratio' after augmentation. Since each subject has 10 original images and augmentation multiplies the data, a random image-level split will likely place images of the same subject in both training and testing. In that case, the high accuracies in Section 5.3 (e.g., CNN 98.404%, Hybrid 100%, LDA+kNN 87.76%) can reflect memorization of identities rather than generalization to unseen people. The paper must specify whether the split was by identity and, if not, re-run the experiments with an identity-disjoint split.
  2. [Section 6] The discussion states that the CNN 'emerged as the most promising in our hypothetical results.' The word 'hypothetical' directly contradicts the empirical framing of Section 5, where the numbers are presented as measured accuracy. The paper needs to clarify whether the results are actual experimental outcomes or illustrative/hypothetical; if they are hypothetical, the central claims of the paper are not supported by data.
  3. [Section 5 and Figure 7] All accuracy, precision, recall, and F1 values are reported as single-run point estimates with no variance, confidence intervals, or number of repeated trials. For small datasets (40 and 50 subjects) and high-capacity models, single-run accuracy is not a reliable basis for the comparative ranking claimed in Sections 5.2 and 5.3. The paper should report multiple runs, cross-validation, or at least error bars, and should make the raw numeric results available as a text table rather than only as the image in Figure 7.
  4. [Section 3.2 and Section 4.4] The preprocessing descriptions are incomplete and somewhat inconsistent. Section 3.2 states that JFAD images are resized to 250x250 for comparison, while Section 3.1 says the LFW subset images are 256x256. Section 4.4 reports that after preprocessing JFAD has 2816 images and LFW has 3508 images, but the augmentation parameters (number of augmented images per original, types and ranges of transformations) are not specified. Without these details, the experimental setup cannot be reproduced, which is especially important because the JFAD is claimed to be a new benchmark resource.
minor comments (4)
  1. [Section 4.1.1] There is a typo in 'correspoinding vector' and inconsistent spelling between 'Minimisation' and the American spellings used elsewhere; these should be corrected.
  2. [References] Several references appear unrelated to face recognition or dataset diversity, including [1] on ambient assisted living, [8] on COVID/vaccine discourse, and [10] on video action recognition. The paper should cite relevant fairness and face-recognition-benchmark literature instead.
  3. [Section 3.2] The JFAD dataset is released only as a Google Drive link with no license, no consent or IRB statement, and no demographic metadata beyond age range. For a dataset intended to serve as a benchmark, these details are important and should be added.
  4. [Section 4.4] The use of RetinaFace to remove images where no face is detected is described without quantifying how many images were removed or whether removal rates differed across demographics; this should be reported to assess potential bias in the preprocessing step.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported accuracies are empirical benchmark results, not derivations that reduce to their own inputs.

full rationale

The paper's central claims are comparative test accuracies of face recognition models on an LFW subset and the newly collected JFAD dataset. These are empirical measurements produced by training and testing pipelines, not mathematical derivations from fitted parameters or prior results. No parameter is fitted to the reported accuracy and then renamed as a prediction; no uniqueness theorem from the authors' prior work is invoked to force a choice; and no stated equation makes a claimed output equal to an input by construction. The cited references are standard external sources (Belhumeur, Deng, LeCun, Szegedy, etc.) and are not self-citations carrying the argument. The one methodological ambiguity is Section 4.4, which says only that 'The datasets were then split into training and testing sets with a 80:20 ratio' without stating that the split is identity-disjoint; this is a generalization-validity and soundness concern, not a circularity concern, because the model outputs are not defined in terms of the reported accuracies and the test performance could in principle differ. Similarly, the phrase 'hypothetical results' in Section 6 weakens the empirical framing but does not constitute a circular reduction. The derivation chain is therefore self-contained with respect to the circularity criteria, and the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claims depend on several unstated or weakly supported modeling and dataset assumptions: the representativeness of a 40-subject dataset, the validity of the augmentation and split protocol, and the correctness of standard subspace and CNN methods. No free parameter values are reported, and the only new entity is the dataset.

free parameters (5)
  • k in k-NN
    The value of k for the PCA, LDA, and Gabor k-NN classifiers is not reported, and accuracy depends on it.
  • Number of PCA/LDA components
    The dimensionality of the projected subspace is chosen by hand and not stated, affecting the reported accuracies.
  • CNN architecture hyperparameters
    The number of filters, learning rate, optimizer, batch size, and epochs are not specified, despite the CNN being a central model.
  • Gabor filter bank parameters
    Frequencies, orientations, and kernel sizes for the Gabor filters are not given, though the Gabor+kNN result is highlighted.
  • Augmentation parameters
    Ranges for shear, rotation, contrast, brightness, and sharpness are not specified, though augmentation determines the final dataset sizes.
assumptions (4)
  • standard math Standard subspace methods (PCA, LDA, LPP) and CNN operations are mathematically correct as applied.
    The paper relies on established projection and convolution techniques without deriving them, which is acceptable but is a background assumption.
  • domain assumption JFAD with 40 subjects from IIT Jodhpur is representative of India's ethnic diversity.
    Section 3.2 claims a 'rich tapestry' of Indian faces, but provides no demographic breakdown, regional distribution, or skin-tone metadata.
  • domain assumption The 80:20 train/test split yields independent training and test sets.
    Section 4.4 states only the split ratio; if it is not identity-disjoint, the same subject appears in both sets and the accuracy numbers are inflated.
  • ad hoc to paper Removing images where RetinaFace detects no face does not bias the dataset.
    Section 4.4 removes undetected faces without analyzing whether detection failure correlates with pose, lighting, or skin tone, which could bias the cleaned dataset.
invented entities (1)
  • IITJ Faces of Academia Dataset (JFAD) independent evidence
    purpose: Serves as the central benchmark for evaluating face recognition models on Indian student faces.
    A public Google Drive link is provided, so the dataset can be inspected, but it lacks demographic metadata, a license, versioning, and external audit. The claim that it represents Indian diversity is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Surveying Facial Recognition Models for Diverse Indian Demographics: A Comparative Analysis on LFW and Custom Dataset." pith.science (2026). https://pith.science/paper/BIKNZJNC

@misc{pith2026241208048,
  author       = {Pith},
  title        = {Pith review of: Surveying Facial Recognition Models for Diverse Indian Demographics: A Comparative Analysis on LFW and Custom Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIKNZJNC}},
  note         = {Machine review of arXiv:2412.08048}
}
read the original abstract

Facial recognition technology has made significant advances, yet its effectiveness across diverse ethnic backgrounds, particularly in specific Indian demographics, is less explored. This paper presents a detailed evaluation of both traditional and deep learning-based facial recognition models using the established LFW dataset and our newly developed IITJ Faces of Academia Dataset (JFAD), which comprises images of students from IIT Jodhpur. This unique dataset is designed to reflect the ethnic diversity of India, providing a critical test bed for assessing model performance in a focused academic environment. We analyze models ranging from holistic approaches like Eigenfaces and SIFT to advanced hybrid models that integrate CNNs with Gabor filters, Laplacian transforms, and segmentation techniques. Our findings reveal significant insights into the models' ability to adapt to the ethnic variability within Indian demographics and suggest modifications to enhance accuracy and inclusivity in real-world applications. The JFAD not only serves as a valuable resource for further research but also highlights the need for developing facial recognition systems that perform equitably across diverse populations.

Figures

Figures reproduced from arXiv: 2412.08048 by the authors.

Figure 1
Figure 1. Methodology overview. 2 Problem Statement The efficacy of facial recognition systems is largely contingent upon the diversity embodied in the training datasets. Predominant datasets like the Labeled Faces in the Wild (LFW) fail to adequately capture the vast spec￾trum of the global population, especially the underrep￾resented ethnic groups from the Indian subcontinent [1]. This deficiency in representation can intro… view at source ↗
Figure 2
Figure 2. Subjects and Images from proposed JFAD dataset (after Face Extraction ) The JFAD’s emphasis on real-world imaging conditions and a focused demographic aims to provide a rigorous testbed for assessing how well facial recognition tech￾nologies can adapt to the ethnic diversity and youthful de￾mographic of modern India. This dataset not only serves as a tool for technological evaluation but also highlights the importan… view at source ↗
Figure 4
Figure 4. Laplacianfaces obtained from LFW dataset [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Feature extraction using Gabor Filter bank [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 8
Figure 8. Figure 8: Accuracy and Loss per Epoch graphs for CNN. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 7
Figure 7. Figure 7: Results table These results suggest that the CNN and Hybrid mod￾els yields high accuracy across both datasets. On the other hand, the traditional LDA+kNN and Gabor mod￾els also showed a marked improvement (with image res￾olution reduced for effective computation ). The…
Figure 9
Figure 9. Figure 9: Accuracy of Different Models on LFW and Cus [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 10 canonical work pages

  1. [1]

    M. H. Abidi, M. K. Mohammed, et al. Ambient assisted living for enhanced elderly and differently abled care: A novel attention transfer learning-based crossover chimp optimization. Journal of Disability Research, 2024

    work page 2024

  2. [8]

    N. Straton. Computational model of engagement with stigmatized sentiment: Covid and general vac- cine discourse on social media. Network Modeling Analysis in Health Informatics and Bioinformatics , 13(1):1–15, 2024. 6

    work page 2024

  3. [10]

    Zhang, Z

    H. Zhang, Z. Hu, D. Yu, L. Guan, X. Liu, and C. Ma. Multipath attention and adaptive gating network for video action recognition. Neural Processing Letters, 2024. 7

    work page 2024

  4. [2]

    P. N. Belhumeur, J. P. Hespanha, and D. J. Krieg- man. Eigenfaces vs. fisherfaces: Recognition us- ing class specific linear projection. IEEE Transac- tions on Pattern Analysis and Machine Intelligence , 19(7):711–720, 1997

    work page 1997

  5. [3]

    Retinaface: Single-stage dense face localisation in the wild

    Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense face localisation in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2019

    work page 2019

  6. [4]

    Kumar and P

    A. Kumar and P. Bhattacharya. Diversity in facial recognition datasets: A case study in india. Journal of Computer Vision, 101(2):123–135, 2015

    work page 2015

  7. [5]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recog- nition. Proceedings of the IEEE, 86(11):2278–2324, 1998

    work page 1998

  8. [6]

    Liu and H

    C. Liu and H. Wechsler. Gabor feature based clas- sification using the enhanced fisher linear discrimi- nant model for face recognition. IEEE Transactions on Image Processing, 11(4):467–476, 2002

    work page 2002

Show all 10 references
  1. [7]

    C. D. Manning, P. Raghavan, and H. Schütze. Intro- duction to Information Retrieval . Cambridge Uni- versity Press, 2008

    2008

  2. [9]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Du- mitru Erhan, Vincent Vanhoucke, and Andrew Ra- binovich. Going deeper with convolutions. In Pro- ceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR) , ...

    2015

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.