Test & measure accuracy of face count detection

[MeenakshiArunsankar]

Description

Establish ground-truth accuracy metrics for the face count detection system used in proctoring.

The frontend currently uses @tensorflow-models/face-detection (MediaPipe face detector) to flag NO_FACE (0 faces) and MULTIPLE_FACES (>1 face) anomalies. However, no accuracy evaluation exists — the model's confidence scores are discarded, and no test set validates behavior across real exam conditions.

Current Implementation

• Detection: frontend/src/components/ai/FaceDetectorWorker.ts runs MediaPipe face detector with maxFaces: 10
• Anomaly mapping: frontend/src/components/floating-video.tsx lines 599-619 — captures screenshot and POSTs to backend
• Backend logging: backend/src/modules/anomalies/controllers/AnomalyController.ts stores evidence but no confidence scores

Work Required

1. Build a labeled test set (~200–300 frames) covering:

   • 0 faces (empty desk, user moved away)
   • 1 face (normal exam position)
   • 2+ faces (peer present, multiple people in frame)
   • Edge cases: masks, partial occlusion, side angles, poor lighting, virtual camera detection

2. Persist confidence scores — modify anomaly endpoints to include per-face detection confidence from the model

3. Run offline evaluation against the test set:

   • Compute precision/recall/F1 for NO_FACE and MULTIPLE_FACES flags
   • Confusion matrix by face count bucket
   • Break down performance by condition (lighting, angle, etc.)

4. Document results in a test report with:

   • Accuracy metrics by condition
   • False positive/negative rates by scenario
   • Recommendations for threshold tuning (if needed)
   • Decision: is current accuracy acceptable for proctoring, or does model need retraining/replacement?

Acceptance Criteria

• Labeled ground-truth test set (CSV: frame_id, ground_truth_face_count, condition_tags)
• Confidence scores logged in anomaly records
• Offline eval script (Python/Node) computing precision/recall/F1/confusion matrix
• Test report with results, broken down by condition
• Documented decision on accuracy adequacy (with evidence)

Related

• Frontend: frontend/src/components/ai/FaceDetectorWorker.ts
• Frontend: frontend/src/components/floating-video.tsx
• Backend: backend/src/modules/anomalies/controllers/AnomalyController.ts

[JoelJosephPhilip]

Hi! I'd like to take this issue.

Plan

1. Persist confidence scores — update the face-count path so MediaPipe/TF.js per-face confidence is sent with anomaly reports and stored on the backend (not discarded).
2. Labeled test set — build a CSV-backed frame set covering 0 / 1 / 2+ faces plus edge cases (lighting, angle, occlusion, masks), targeting ~200–300 frames (I'll start with a smaller seed set and expand).
3. Offline evaluation — add a script (Python or Node) that runs face-count detection over the test set and reports precision / recall / F1, a confusion matrix, and breakdowns by condition tags.
4. Test report — document metrics, false positive/negative rates by scenario, threshold recommendations, and a clear decision on whether current accuracy is acceptable for proctoring.
   I'll open a PR (or split PRs if cleaner: confidence plumbing first, then eval harness + report) and link it here.
   Please assign this to me if that works. Thanks!