Model reference — Pictures · Video · Live feed
Face recognition matches a face against a list of enrolled people and returns who it is, or reports no match. It is a search, and it is materially harder than confirming that two pictures show the same person.
Every deployment is really three models: one that finds faces, one that turns each face into a numeric fingerprint, and — on video — one that follows a person between frames so the same face is not counted twice. The decision threshold is where the engineering meets the policy: the larger your enrolled list, the more chances of a false match, so the threshold must be set against list size and the cost of an error. This is regulated technology in most jurisdictions, requiring a lawful basis, retention limits and notice, and we will not deploy it without those settled.
Models in this group take a single image as input: a photograph, a scan or a screenshot. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one detected face matched against an enrolled list of 10,000 people.
Vendor: InsightFace
What it does: turns a face into a numeric fingerprint and matches it against the enrolled list. The most widely deployed and best documented option, with published accuracy on standard benchmarks.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Michigan State University
What it does: handles poor-quality captures — small, blurred, badly lit — considerably better, which is where most recognition failures actually occur.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 36,000 | ≈ 140,000 | ≈ 400,000 |
Vendor: Google
What it does: the established earlier model, less accurate but exceptionally well understood, which matters where the system must be explained to a regulator.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 50,000 | ≈ 190,000 | ≈ 520,000 |
Vendor: InsightFace
What it does: finds and straightens faces before matching. Alignment quality affects accuracy more than the choice between fingerprinting models.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 20,000 | ≈ 80,000 | ≈ 220,000 |
Models in this group take a recorded video file as input and are applied frame by frame. The three hardware tiers mean the same as above. The sample input here is one minute of 1080p video at 25 frames per second, with faces recognised and tracked.
Vendor: InsightFace
What it does: finds faces on every frame fast enough to keep up with recorded video at full frame rate, which is the constraint that rules out slower detectors.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 900 | ≈ 3,000 | ≈ 8,000 |
Vendor: Huazhong University of Science and Technology
What it does: follows each person between frames so a face seen for five seconds is one identification rather than a hundred and twenty. This is what makes video recognition affordable and its counts meaningful.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 1,200 | ≈ 4,000 | ≈ 10,000 |
Vendor: Tel Aviv University
What it does: a tracker that re-identifies a person after they are briefly hidden — behind a pillar, another person — which prevents one visit being recorded as two.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (video minutes processed per hour) | ≈ 800 | ≈ 2,800 | ≈ 7,000 |
Models in this group take a live camera or stream as input and must keep pace with it in real time. The three hardware tiers mean the same as above, but the rate here is how many streams one server of that tier can keep up with in real time rather than a per-hour count. The sample input here is one 1080p stream with faces recognised as they appear.
Vendor: Shenzhen Institute of Advanced Technology
What it does: finds faces on a live stream using very little computing power, small enough to run on the camera-side device so that images need not be sent anywhere.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 12 streams | ≈ 40 streams | ≈ 120 streams |
Vendor: InsightFace
What it does: the accurate live detector, fast enough to hold real time across several streams on one card.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 6 streams | ≈ 24 streams | ≈ 70 streams |
Vendor: InsightFace
What it does: matches each detected face against the enrolled list within the frame budget, which is what makes recognition at a door or gate immediate.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 8 streams | ≈ 30 streams | ≈ 90 streams |
Vendor: Huazhong University of Science and Technology
What it does: follows people through the live scene so an alert fires once per person rather than repeatedly while they remain in view.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (camera streams handled at 25 frames per second) | ≈ 10 streams | ≈ 36 streams | ≈ 100 streams |
The right configuration depends on enrolled list size, capture conditions, and the relative cost of a false match against a missed one. Our consultants measure both error rates on your own material, then recommend the detection, fingerprinting and tracking models, the threshold, and the review process for uncertain matches.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Face recognition. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.