Face Recognition
Use on-device ArcFace face recognition to identify and track individuals across video evidence with privacy-first controls and configurable confidence thresholds.
Face Recognition#
FrameCounsel includes on-device face recognition powered by the ArcFace model, enabling defense teams to identify and track individuals across multiple video sources without uploading any data to external servers. This capability is critical for establishing who was present at a scene, tracking witness movements, and verifying officer identities.
ArcFace Model Overview#
ArcFace is a state-of-the-art face recognition model that produces high-dimensional embeddings for each detected face. FrameCounsel runs the model through Core ML, fully optimized for Apple Silicon. Key characteristics:
- 99.8% accuracy on standard benchmarks under good lighting conditions
- Robust to moderate occlusion, low resolution, and non-frontal angles
- Processes approximately 30 faces per second on M2 Pro hardware
- All inference runs locally with zero network activity
No Cloud Processing
Face recognition in FrameCounsel is entirely on-device. Face embeddings are stored only within your encrypted case project. No biometric data is transmitted, stored remotely, or accessible to FrameCounsel as a company.
Adding Reference Photos#
To identify specific individuals in your video evidence, add reference photos to the case:
- Open the Identity Panel from the sidebar or press
- Click Add Person and provide a name and role (e.g., "Officer Smith - Arresting Officer")
- Drag one or more reference photos into the person's profile, or extract a face directly from any video frame by right-clicking a detected face
- FrameCounsel generates an embedding for each reference photo and uses the strongest match for identification
For best results, provide 3-5 reference photos per person showing different angles and lighting conditions.
Matching Across Videos#
Once reference identities are established, FrameCounsel scans all video evidence in the case and identifies appearances of each known individual. Results appear in the Identity Panel with:
- Timestamp and source video for each appearance
- A thumbnail of the matched face alongside the reference photo
- Confidence score for the match
- Duration of presence in the frame
Click any match to jump directly to that moment in the video player.
Confidence Thresholds#
Configure the minimum confidence score required for a positive identification in Settings > Face Recognition > Threshold:
| Threshold | Recommended Use |
|---|---|
| High (0.85+) | Court submissions where certainty is critical |
| Medium (0.70-0.85) | Investigation phase for generating leads |
| Low (0.50-0.70) | Exploratory analysis in degraded footage |
Review All Matches
Automated face recognition can produce false positives, particularly in low-light or low-resolution footage. Every match should be reviewed by a human before being relied upon in legal proceedings.
Privacy Controls#
FrameCounsel provides granular privacy controls for face recognition data:
- Auto-purge - Automatically delete face embeddings when a case is closed
- Scope restriction - Limit recognition to specific evidence items rather than the entire case
- Audit logging - Every recognition query is logged with the user, timestamp, and parameters
- Export controls - Face recognition data can be excluded from report exports
Next Steps#
Extend your tracking capabilities with Object Tracking to follow vehicles, weapons, and other items of interest across your video evidence.