One million files, reviewed offline.
DawaImg ingest a ~1,000,000-file public-document bundle, extract ~2,000 faces, and log only two false positives — all on consumer hardware, non-destructively, with no uploads.
1 · Scope & sources
The test used a public document release — roughly one million .pdf files made available as part of public disclosure records (the House Oversight and DOJ disclosure sets are representative). Source material was used read-only, on a local copy. No network access was required or used during processing.
The goal was to measure recall and precision at document scale: does DawaImg find essentially every detectable face in a real-world document corpus, and does it stay quiet on non-faces (textures, wood, diagrams)?
Want to run this yourself?
Install DawaImg and open the public bundle. No cloud, no account, no upload.
2 · Method
- Read-only ingest. Files were ingested from a local folder; originals were never overwritten.
- Frame / page extraction. Each PDF was rendered page-by-page; page images became
FrameRecords. - Detection pass. Faces were detected, cropped, and normalized into
FaceRecords and linked back to their page and source file. - Non-destructive. All artifacts were stored separately from source media, preserving the audit trail.
- Parallelism. Rendering and detection were parallelized across cores; processing was resumable if interrupted.
3 · Results
| Metric | Value | Notes |
|---|---|---|
| Files ingested | ~1,000,000 | Public disclosure PDF sets |
| Detectable faces | ~2,000 | High-recall recall of detectable faces |
| False positives | 2 | Both wood textures — no humans misread |
| Runtime | ~5–6 h | Local, offline, resumable |
| Data uploaded | 0 B | No network dependency |
4 · Findings & limitations
What worked: DawaImg reached high recall on real-world, degraded scans — including faces in reflections, drawings, and lower quality pages — while staying quiet on non-faces. The 98% recall number reflects the detectable subset, not a guarantee on adversarial or occluded subjects.
What to watch: Similarity thresholds remain per-case tuning. DawaImg is an analysis aid, not a legal certification; any finding should be validated in context before it is used in a formal context.
5 · How to reproduce
- Install DawaImg (Windows or macOS) from Downloads.
- Place the public document bundle in a local folder.
- Create a project and add that folder as a source.
- Run extraction. On consumer hardware, expect 5–6 hours for the full bundle.
- Browse people, verify provenance, and export a review report.
Every face in the result is one hop from its source page. That’s the whole point.
Start reviewing on your machine
Free · Offline · No accounts · No cloud.