Towards Robust Classroom Attendance: A Comprehensive Evaluation of Face Detection and Recognition Models
Himani Trivedi, Hiren Patel, Ridham Patel, Krutika Patel, Nancy Patel
cs.CV · cs.LG
Quick verdict
A classroom face dataset + YOLOv8n/FR evaluation; five models tie at 99.75% Top-1 on 400 test images, so the FaceLiVTv2-M 'best' claim is weakly supported.
Summary
The paper introduces the Visage Face Dataset (VFD), a two-part classroom collection — VFD-Detection (801 hand-captured classroom photos; Table I lists 3,635 images and 96,409 face annotations after augmentation) and VFD-Recognition (3,500 aligned 112x112 crops of 100 student identities) — and evaluates a two-stage attendance pipeline. From a YOLOv8 family comparison it selects YOLOv8n (mAP@0.5 0.935, F1 0.91, 3.2M parameters), then fine-tunes 13 configurations of seven recent recognition models (LVFace, QCFace, FaceLiVTv2, TopoFR, EdgeFace, TransFace, GhostFaceNets) with softmax cross-entropy and classifies test faces by cosine similarity to gallery centroids. Five configurations tie at 99.75% Top-1 on the 400-image test set; the abstract singles out FaceLiVTv2-M (99.75% Top-1/Top-5, 6.459 ms on a T4), yet EdgeFace-Base matches that Top-1 at 4.029 ms and three models reach 100% Top-5. Main limitations: augmentations were applied within the test split, creating leakage risk; 400 test images make rank differences equal to single samples, with no significance testing; the claimed 12-detector comparison is not tabulated; dataset accounting (16,234 samples vs Table I) is unreconciled; and no public code or data release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The Visage Face Dataset, claimed at 16,234 face samples, is a realistic and challenging benchmark for face detection and recognition in classroom attendance.
- YOLOv8n is the best detection choice for the pipeline, delivering mAP@0.5 of 0.935 and F1 of 0.91 with only 3.2M parameters.
- FaceLiVTv2-M achieves the best recognition result with 99.75% Top-1/Top-5 accuracy at 6.459 ms inference on a Tesla T4 with 7.02M parameters.
- Small or hybrid recognition models match much larger ones, making real-time classroom attendance feasible on low-resource devices.
Strengths
Realistic, ethically cleared data collection
801 classroom photos hand-captured across schools with pose, illumination, expression and partial occlusion variation (§III-A); explicit consent statement and team-only access policy (§III-B 'Ethics and Privacy').
Unified protocol across very recent FR architectures
13 configurations of 7 models published 2023-2026 (LVFace ICCV'25, TopoFR NeurIPS'24, QCFace WACV'26, FaceLiVTv2, EdgeFace, TransFace, GhostFaceNets) share identical alignment (RetinaFace + ArcFace similarity to 112x112), identical fine-tuning (§IV-B) and identical centroid evaluation — a like-for-like comparison rarely attempted.
Careful latency measurement
Explicit efficiency trade-off reasoning for detector choice
Weaknesses
Headline 'best model' claim contradicted by its own Table III
Augmentation inside the test split creates leakage risk
Test set too small to rank models; no statistical analysis
Claimed experiments and dataset numbers do not reconcile
No public artifacts and motivation axes never measured
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.22750, cs.CV) contributes the Visage Face Dataset (VFD) for classroom attendance — split into VFD-Detection and VFD-Recognition — plus a two-stage detection+recognition evaluation. Four headline claims (§I): (1) VFD is a realistic, challenging classroom benchmark; (2) YOLOv8n is the best accuracy/efficiency detector (mAP@0.5 0.935, F1 0.91, 3.2M params); (3) 13 configurations of seven recent FR models were evaluated under one shared protocol; (4) FaceLiVTv2-M is the best recognizer with 99.75% Top-1/Top-5 at 6.459 ms on a T4. The evaluation design is sound in principle; claim (4) is not supported by the paper's own Table III, and claim (1) rests on numbers that never reconcile.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Visage Face Dataset and how big is it?
It is a classroom face dataset introduced by this paper for attendance research. It has two parts: VFD-Detection (801 hand-captured classroom photos, listed as 3,635 images with 96,409 face boxes after augmentation) and VFD-Recognition (3,500 aligned face crops of 100 students, resized to 112x112). The abstract claims 16,234 total face samples, a figure the paper's own tables do not clearly reconcile.
Which face recognition model does the paper say is best for classroom attendance?
The abstract names FaceLiVTv2-M (99.75% Top-1/Top-5, 6.459 ms on an NVIDIA T4, 7.02M parameters). However, four other configurations (FaceLiVTv2-L, EdgeFace-Base, LVFace, TransFace) tie at the same 99.75% Top-1 on the 400-image test set, and EdgeFace-Base is faster (4.029 ms), so the 'best' designation is not clearly supported by the results table.
Is the Visage Face Dataset or the code publicly available?
As published, no. The paper provides no download link or repository; the ethics statement says consent was obtained and access is limited to the research team. No code or model weights are mentioned.
What detection model does the attendance system use?
YOLOv8n, the nano variant of YOLOv8, chosen for its 3.2M parameters. It reaches mAP@0.5 of 0.935 and an F1-score of 0.91 (precision 0.932, recall 0.886) on the paper's classroom detection dataset, trading a little recall for a 21x smaller model than YOLOv8x.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 (sample images, real-time detection examples, comparison graphs) cannot be analyzed from extracted text; some claimed comparisons (e.g., the 12-detector evaluation) may exist only inside figures.
- The extracted text contains minor OCR artifacts ('attendance,e', stray commas), but all cited numeric values in Tables I-III appear intact and were cross-checked for internal consistency.
- No supplementary material (per-model hyperparameters, augmentation magnitudes, exact epoch counts) is present, limiting judgment on training fairness across the 13 configurations.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.