OUR AI MODELS
Benchmark Methodology
Every accuracy figure published on this site, with the dataset, metric, operating point, and date behind it. If a number is not on this page, we do not publish it.
How to read a row. An accuracy figure quoted without a false-match rate cannot be checked, and cannot be compared to anyone else’s. Every figure below names the corpus it was measured on, how many comparisons it came from, and the threshold it was measured at, so you can reproduce it or hold us to it.
AI MODEL
Face Verification 1:1 (Server Model)
| Claim | 99.45% ± 0.18% verification accuracy (LFW, 10-fold, end-to-end) 94.41% true accept at a 1-in-10,000 false-match rate (VGGFace2) |
| Dataset | LFW: 13,222 images, 5,743 identities VGGFace2: 196,324 images, 540 identities |
| Task | 1:1 verification |
| Metric | 10-fold accuracy, official LFW protocol (threshold refitted per fold) True accept rate at a fixed false-match rate |
| Operating point | Access Control: cosine ≤ 0.6995, FMR 1 in 10,000 Standard: cosine ≤ 0.8275, FMR 1 in 100 |
| Comparisons behind it | 87,161,944 impostor pairs (LFW) · 19,232,832,135 (VGGFace2) Every image against every other image, not a 6,000-pair sample |
| Equal error rate | 0.46% (LFW) · 2.88% (VGGFace2) |
| Detector in the loop | Yes. Figures are end-to-end, from full photograph to decision |
| Hardware | Intel Core Ultra 7 155H, 16 GB DDR5 RAM |
| Model version | AR-FaceEmb-v20-S |
| Date tested | 29 Aug 26 |
AI MODEL
Face Verification 1:1 (Edge Model)
| Claim | 99.18% ± 0.19% verification accuracy (LFW, 10-fold, end-to-end) 85.40% true accept at a 1-in-10,000 false-match rate (VGGFace2) |
| Built for | On-camera and edge deployment. 13 MB model, no GPU required |
| Dataset | LFW: 13,222 images, 5,743 identities VGGFace2: 196,324 images, 540 identities |
| Task | 1:1 verification |
| Metric | 10-fold accuracy, official LFW protocol True accept rate at a fixed false-match rate |
| Operating point | Access Control: cosine ≤ 0.6875, FMR 1 in 10,000 Standard: cosine ≤ 0.8185, FMR 1 in 100 |
| Equal error rate | 0.49% (LFW) · 3.51% (VGGFace2) |
| Cost against Server | 85.40% against 94.41% true accept at the same false-match rate Same security, more false rejections: roughly 1 in 7 legitimate attempts against 1 in 18 |
| Hardware | Intel Core Ultra 7 155H, 16 GB DDR5 RAM |
| Model version | AR-FaceEmb-v20-E |
| Date tested | 29 Aug 26 |
AI MODEL
Face Identification 1:N
| Claim | 99.38% rank-1 against a gallery of 1,260 enrolled people 95.68% rank-1 against a gallery of 405, on markedly harder imagery |
| Open-set claim | 93.71% identified correctly while wrongly flagging 1% of strangers The realistic access-control case: the person may not be enrolled at all |
| Gallery size | 1,260 enrolled identities (LFW) · 405 (VGGFace2) A 1:N figure without a gallery size means nothing, so ours is stated |
| Dataset | LFW: 6,126 probes, 5,836 non-enrolled VGGFace2: 147,537 probes, 48,382 non-enrolled |
| Task | 1:N identification, closed set and open set |
| Metric | Closed-set rank-1 (CMC) Open-set true positive identification rate at fixed false positive identification rate |
| Operating point | Access Control 1:N: cosine ≤ 0.6390, 1% of strangers flagged, gallery of 1,000 1:N uses a tighter threshold than 1:1, because every query is compared against every enrolled person |
| Also within top 5 / top 10 | 99.46% / 99.48% (LFW) · 96.52% / 96.89% (VGGFace2) |
| Hardware | Intel Core Ultra 7 155H, 16 GB DDR5 RAM |
| Model version | AR-FaceEmb-v20-S |
| Date tested | 29 Aug 26 |
Where face recognition usually breaks
Hard Cases
People enrol once and are recognised at an angle, years later. These are the four benchmarks that measure that, and they are the ones worth reading.
| CONDITION | SERVER MODEL | EDGE MODEL |
|---|---|---|
| Cross-pose (CPLFW) | 94.58% | 92.47% |
| 30-year age gap (AgeDB-30) | 98.43% | 96.60% |
| Frontal against profile (CFP-FP) | 97.74% | 95.60% |
| Cross-age (CALFW) | 96.15% | 95.52% |
| Frontal against frontal (CFP-FF) | 99.90% | 99.67% |
AI MODEL
Person and Vehicle Detection
| Claim | 93.4 mAP@0.5 |
| Dataset | VisDrone |
| Classes | person, vehicle |
| Metric | mean Average Precision at IoU 0.5 |
| Hardware | Intel Core Ultra 7 155H, 16GB DDR5 RAM |
| Model version | AR-GenModel-v16 |
| Date tested | 25 Sept 25 |
Test It On Your Own Footage
A feasibility assessment runs our models against your cameras and your conditions, and reports what accuracy you would actually get. Test fees deducted from the project fee if you proceed.