Reference
How Sherlock AI Face Search works
What face search is
Face search takes one photo of a face and finds other photos of the same face elsewhere, even when the pose, lighting, and background differ. It is not the same as Google reverse image search, which looks for copies of the whole image. Sherlock compares the face across different photographs.
The four stages, in detail
The pipeline is a standard, well-established sequence. All four stages run in your browser; only the final 512-number faceprint is sent to our servers.
1. Detect. A compact SCRFD detector (about 2.5 MB) finds each face and five landmarks — both eyes, the nose tip, and the two mouth corners.
2. Align. Using those five points, the crop is warped with a similarity transform to a canonical 112×112 frame. Alignment is what lets a tilted photo match a straight-on one; skipping it collapses accuracy.
3. Embed. A MobileFaceNet ArcFace model turns the aligned crop into a 512-dimension vector — the faceprint — which is then L2-normalized to unit length so that similarity scores stay comparable.
4. Search. The faceprint is compared against the index using cosine similarity, then ranked and thresholded.
Similarity thresholds
Scores are cosine similarity on the normalized 512-dimension vectors:
- 0.65 and above — strong match, shown prominently.
- 0.50 to 0.65 — possible match, shown and labeled.
- 0.40 to 0.50 — weak, shown only behind a “show weak matches” control.
- Below 0.40 — discarded and never returned.
What it cannot do
Sherlock returns images and the pages that host them. It does not return names, addresses, phone numbers, employers, or any assembled profile. A match is a similarity signal, not an identity. Turning a face into a dossier is out of scope by design.
Accuracy and false positives
On clean, well-lit photos, modern face-recognition models exceed 99% accuracy in NIST FRVT benchmarks. Real-world results are lower. Siblings, twins, and unrelated lookalikes return high scores. The embedding approach is the ArcFace method described in Deng et al., 2019. Treat every match as a lead requiring independent confirmation.
How this differs from Google reverse image search
Reverse image search finds copies and near-copies of the same image file. Face search matches the same person across different images — different clothes, different day, different camera. They answer different questions.
Law and biometric data
Face vectors are biometric data. In the EU they fall under GDPR Article 9. In Illinois, the Biometric Information Privacy Act (BIPA) carries a private right of action. We block access from Illinois and Texas until we meet those requirements.
How to get removed
Submit a request at /remove. No account is needed. We generate a faceprint, add it to our exclusion list, and stop returning matches for that face within 30 days.
Frequently asked questions
- How much does Sherlock AI Face Search cost?
- Searches run on credits. A pack of 15 is $4.99 and never expires; monthly plans start at $14.99. Every search returns full source links.
- Does it work on Android?
- Yes. Sherlock runs in any modern mobile browser and installs to your home screen. A Play Store version is coming.
- Is face search legal?
- In most places, searching publicly posted images is legal for personal use. Illinois and Texas have biometric laws we currently block for. Using results to stalk or harass anyone is prohibited and will get the account terminated.
- Do you store my photo?
- No. Analysis happens in your browser and only the faceprint is sent, then discarded.
- Will the person know I searched for them?
- No. Searches are not disclosed to anyone.
- Why did it find nothing?
- Most likely the face is not in the index. Coverage depends on what is publicly posted, and someone with a small online presence may return nothing at all. Blurry photos, extreme angles, and small faces also reduce matches.
- How accurate is it?
- Modern face models exceed 99% on clean, well-lit photos in NIST benchmarks. Real-world results are lower. Treat every match as a lead.
- How do I remove myself?
- Submit a request at /remove. No account needed.