Machine learning · 2026
13k Vision
Study the whole identification workflow, from capture to review.
Reliable identification at a supervised station requires more than a candidate match. This project combines the application with experiments around image quality, gallery size, uncertain matches and the consequences of an incorrect identity. The prototype has been tested on public datasets and simulated station activity; validation in the intended setting is still ahead.
- Status
- Research prototype
- Domain
- Computer vision · Statistical evaluation
- Role
- Evaluation design, local inference, application workflow & research
- Year
- 2026
Explore the workflow with illustrative content.
Start with a sample observation
An abstract mesh represents the input observation. It is a geometric illustration, not a person or biometric record.
Compare candidate patterns
The candidate layer shows a clear or ambiguous example. The bars are illustrative relative similarities, not measured model scores.
Keep uncertainty visible
A clear example and an ambiguous example lead to different review states. This demonstrates the decision boundary without performing recognition.
Illustrative content · no live data
Explore how a candidate reaches a review decision
Stack
- Python
- FastAPI
- ONNX Runtime
- PyTorch
- OpenCV
- SQLite
- JavaScript
A local face-identification research system with a supervised capture interface, human-review paths and a benchmark harness for testing difficult conditions.
Build the review path alongside the model
The application supports enrolment, capture, matching and supervisor review.
A result can be accepted, sent for confirmation or returned for a retake or manual fallback. The interface shows the decision and its reason, making uncertainty part of the workflow.
Test the conditions around recognition
The harness varies image quality, gallery composition, enrolment images and recognition backends.
It records failures to detect a face as well as failures to identify one. A station emulator then exercises the server over repeated transactions, exposing behaviour that an isolated model benchmark cannot show.
Check the benchmark itself
Combining public face datasets introduced an important problem: the same person could appear under different labels.
That can distort both the matching result and the threshold used to judge it. Identity audits and cached embeddings made it possible to investigate the labels and rescore affected runs without repeating all the inference work.
Learn from the experiments that did not help
More elaborate models were not consistently better.
The research includes comparisons of model combinations, decision rules and adaptation, keeping negative findings alongside improvements. Capture quality and the evaluation protocol proved as important to investigate as the recogniser itself.
September 2026: prototype and measured research
The work grew from a design paper into a local service, browser interface, streaming benchmark and station emulator.
September runs added public-dataset stress tests and an acceptance study. The paper remains a working draft, and those results have not established performance on images from the intended deployment.
Pipeline
Supervised identification prototype
Stage Can reject
-
01
Capture
Cooperative face images
-
02
Check quality
Usable frames or retake
-
03
Local inference
Face embeddings
-
04
Search gallery
Candidate scores
-
05
Apply decision rule
Accept, review or fallback
-
06
Confirm and record
Reviewed transaction
Behind the work
The skills
that came together.
Evaluation
- One-to-many identification protocols
- Paired comparisons and uncertainty intervals
- Label-noise audits and reproducible rescoring
- Image-quality and degradation experiments
Implementation
- Local GPU inference
- Streaming gallery evaluation
- Supervised capture and review interfaces
- Station emulation and transaction logs
Project highlights
-
A complete prototype workflow
Enrolment, capture, matching and human review
-
Evaluation beyond a headline score
Detection, identification and decision outcomes kept separate
-
Reproducible experiments
Cached inference, label audits and traceable rescoring
-
Limits stated clearly
Public-data experiments, with site validation still pending