Kawal511/sentinelvision-proctoring-yolov8n

SentinelVision — YOLOv8n fine-tuned for exam proctoring

A YOLOv8-nano detector fine-tuned to spot the objects that matter in an online exam: phones, books, headphones, laptops, TVs and people.

Used by SentinelVision, a real-time webcam proctoring app.

Why it exists

Stock COCO YOLOv8n is poor at this task. Phones in exam footage are small, dim, hand-occluded and shot through a cheap webcam — nothing like the clean, centred phones in COCO. Measured on held-out proctoring images, the pre-trained model reaches an F1 of 0.203, and no confidence threshold rescues it. That is a model problem, not a tuning problem, which is what motivated fine-tuning.

Results

Both models were scored by the same from-scratch AP/F1 implementation on the same held-out images. The two number their classes differently (COCO class 67 vs class 1 here), so neither model's self-reported metric is comparable to the other's.

Model	Precision	Recall	F1
Pre-trained YOLOv8n (COCO)	0.215	0.192	0.203
This model	0.910	0.944	0.927

Measured at confidence 0.35, chosen by sweeping 0.05–0.95 against the full 1,822-image held-out split. The F1 curve is flat from 0.25–0.50, so that is a plateau rather than a knife edge. 64% of those images contain no phone at all, so precision is measured against genuine opportunities to hallucinate.

Classes

0 book   1 cell phone   2 headphone   3 laptop   4 person   5 tv

Resolve classes by name, not by index. This model's cell phone is class 1; in COCO it is 67. A hard-coded index fails silently — it matches nothing and looks like a detector that found nothing, rather than a lookup that was wrong.

Training data

Online-proctoring-system via Roboflow Universe — 25,173 webcam-framed exam images, CC BY 4.0, 6 classes.

The dataset's own split is unusable: one source video is ~87% of the data and supplies 100% of validation and test. Training on it and validating on frames of the same person in the same room reports a number that says nothing about anyone else. The split was regrouped by source video, so validation is people the model never trained on.

Fine-tuned on a Kaggle GPU from yolov8n.pt, with the backbone frozen — with a dataset this size, updating all layers mostly memorises, and the early layers already encode generic edges and textures worth keeping.

Intended use

Educational and research work on automated proctoring, and as a worked example of fine-tuning a detector for a specific visual domain.

Not suitable for making decisions about real people. See the limitations below: on an unfamiliar camera this model raises frequent false alarms, and a false accusation of cheating is a serious harm. Any real deployment would need explicit consent, a human in the loop, and validation in the actual deployment environment.

Limitations — measured, not guessed

Precision does not transfer to an unseen camera. On 741 phone-free frames from a webcam the model had never seen, it fired cell phone on 56.4% of them. Every false positive was the same object: a wall shelf holding a keyboard. Live it scored that shelf 0.60–0.71, against a real phone at 0.72–0.79 — the ranges touch, so no confidence threshold separates them. A blank-background control confirmed the shelf was the sole source.

The 0.910 precision above is a statement about the public validation set, not about any particular room. This is textbook domain shift, and it is the single most important thing to know before using these weights.

Partially visible phones are missed. Held low and half out of frame, phones score only 0.09–0.16 — below any usable threshold.

Mitigation used downstream. SentinelVision does not solve this with a threshold. It suppresses detections in regions of the frame that never move, since a shelf is bolted to the wall and a phone in use is not — with an exemption for anything overlapping the detected person, because a phone being read is held still. The honest fix remains hard negatives collected from the deployment environment.

Usage

from ultralytics import YOLO

model = YOLO("proctoring_yolov8n_best.pt")

# Look the class up by NAME -- never hard-code the index.
phone = next(i for i, n in model.names.items() if n == "cell phone")

results = model("frame.jpg", classes=[phone], conf=0.35)

Use conf=0.35 to reproduce the reported metrics. In a live room a higher floor (~0.60) plus static-region suppression is what makes it usable — see the repository.

Licence

AGPL-3.0, inherited from Ultralytics YOLOv8, which these weights are derived from. If you use this in a network service, AGPL obligations apply.

The training data is CC BY 4.0 and requires attribution to the Online Proctoring System dataset (online-exam-cheating-detection workspace, Roboflow Universe).

The thresholds used alongside this model were calibrated against two further public datasets, credited in full in the repository's dataset notes: the Gourier Head Pose Image Database (Gourier, Hall & Crowley, 2004; CC BY 4.0 as redistributed) and MichalMlodawski/closed-open-eyes (ODC-BY).

Citation

@software{sentinelvision_proctoring_yolov8n,
  title  = {SentinelVision: YOLOv8n fine-tuned for exam proctoring},
  url    = {https://github.com/kcosteen/SentinelVision},
  year   = {2026}
}