Missing Persons Finder

Face-recognition admin panel for registering missing-person reports and matching new photographs against them using vector similarity search.

Built with InsightFace embeddings, ChromaDB vector search, and a PyQt5 desktop interface.

Overview

Searching a missing-persons registry by name fails exactly when it matters most — when the person is found but cannot identify themselves. This tool searches by face instead.

Every registered post is reduced to a 512-dimensional face embedding and stored in a persistent vector index. Matching a new photo becomes a nearest- neighbour lookup that stays fast as the registry grows, rather than a linear scan over every stored image.

Key features

Face embedding extraction

• InsightFace buffalo_l model
• Automatic CUDA execution with transparent CPU fallback
• 512-dimensional embedding per detected face
• Model loaded once behind a thread lock and reused

Vector search with ChromaDB

• Persistent client, so the index survives restarts
• HNSW indexing with cosine distance
• Add, delete, and similarity-search operations
• Rebuilds itself automatically from posts.json when the collection is empty

Admin dashboard (PyQt5)

• Post feed with auto-refresh
• Add-post form with multi-image upload
• Search view and dedicated results view
• Full-size image viewer

Architecture

flowchart LR
    A[Photograph] --> B[InsightFace buffalo_l]
    B --> C[512-d embedding]
    C --> D[(ChromaDB<br/>HNSW · cosine)]
    Q[Query photo] --> B
    D --> E[Ranked matches]
    E --> F[PyQt5 results view]
    G[(posts.json)] -.->|rebuild when empty| D

Loading

posts.json is the source of truth for post metadata; ChromaDB is a derived index that can always be rebuilt from it.

Project structure

main.py              Application entry point and window wiring
config.py            Paths, thresholds, and runtime settings
face_model.py        InsightFace loading and embedding extraction
chroma_manager.py    Vector store lifecycle, query, and rebuild
utils.py             Post persistence and cosine similarity
ui/
  feed_widget.py           Post feed with auto-refresh
  add_post_widget.py       New-post form and image upload
  search_widget.py         Search input and controls
  search_results_widget.py Ranked match display
  image_viewer.py          Full-size image viewer

Configuration

Defaults live in config.py:

Setting	Default	Meaning
SIMILARITY_THRESHOLD	0.20	Minimum cosine similarity to report a match
OUTLIER_HIGH_THRESHOLD	0.85	Upper bound used in outlier checks
OUTLIER_LOW_THRESHOLD	0.25	Lower bound used in outlier checks
MAX_IMAGES	5	Images allowed per post
AUTO_REFRESH_MS	3000	Feed refresh interval

Post images are written to posts/ and metadata to posts.json, both resolved relative to the application directory. The same resolution works when the app is frozen into a standalone executable.

Installation

pip install -r requirements.txt

For GPU inference, replace onnxruntime with onnxruntime-gpu and install a matching CUDA runtime. The application falls back to CPU automatically if the CUDA provider cannot be initialised, so a GPU is optional.

Usage

python main.py

InsightFace downloads the buffalo_l model weights on first run.

Limitations

• Similarity thresholds are fixed constants, not calibrated against a labelled evaluation set — treat ranked results as candidates for human review, never as identification.
• Recognition accuracy degrades with pose, occlusion, low resolution, and large age gaps between the registered and query photographs.
• The admin panel has no authentication layer; it assumes a trusted operator on a trusted machine.
• All data is stored locally in plain files. Deploying this against real missing-persons data would require access control, encryption at rest, and a retention policy.

License

MIT