01
Single frames contain misleading patterns
Circular objects, paired highlights and high-contrast textures can resemble a face to a detector. A system that immediately blurs every candidate can turn brief uncertainty into a visible distraction.
Temporal confirmation asks whether a candidate behaves like a coherent subject across neighbouring frames. This filters many unstable detections before they become tracks.
02
Thresholds create a real trade-off
Raising a confidence threshold can reduce false positives but may also miss small or partial faces. Lowering it can increase coverage and review work. There is no universal setting for every camera and risk level.
- Use representative footage for calibration.
- Choose a stricter privacy mode for high-risk work.
- Review both false positives and false negatives.
03
Corrections should stay project-specific
Removing a false track should correct the current project, not silently retrain a general model from sensitive footage. Project-local corrections keep the action predictable and avoid unintended learning from private video.
Track recurring conditions—such as a specific camera or night scene—so future jobs start with better review assumptions.
Apply the workflow
Review real footage locally.
CensorFlow detects and tracks visible faces on your Windows PC, then lets a person correct and approve the result before export.