PartiPix

Photo sharing platform with facial recognition.

GoPythonHTML/CSS

Basic Idea

At events, photos get scattered across everyone’s phones. Partipix centralizes them: guests join an event via a share code, upload their photos, and the platform automatically surfaces every photo each person appears in using facial recognition.

Architecture

The backend is a Go service using fasthttp for HTTP handling and SQLite via GORM for persistence. The frontend is a SvelteKit app. When a photo is uploaded, it is passed synchronously through the facial recognition pipeline before the API response is returned, so the client receives face match data alongside the freshly saved photo.

Upload Photo
     │
     ▼
Extract faces (dlib via go-face)
     │
     ▼
Compute 128-dim descriptor per face
     │
     ▼
Compare against stored embeddings (event participants only)
     │
     ▼
Persist: PhotoPersonMatch + FaceLocation
     │
     ▼
Return Photo + []FaceMatch to client

Facial Recognition Pipeline

Face detection and embedding are both handled by github.com/Kagami/go-face, which wraps dlib’s ResNet-based face recognition model (dlib_face_recognition_resnet_model_v1.dat). Landmark detection uses shape_predictor_5_face_landmarks.dat and the face detector uses mmod_human_face_detector.dat.

Reference photos: each user uploads one or more profile photos. On upload, the backend detects the face, extracts its 128-float32 descriptor, serializes it to 512 bytes (float32 × 128), and stores it in the face_embeddings table. This pre-computation means matching at photo-upload time is a pure in-memory vector comparison loop rather than running the full recognition model again.

Matching: for each face detected in an event photo, the backend:

  1. Computes the face descriptor from the cropped region (padded 20% on each side, resized to 150×150 with Lanczos).
  2. Loads stored embeddings only for participants in that event.
  3. Computes squared Euclidean distance between the query descriptor and each stored embedding; similarity = 1.0 − distance.
  4. Records a match if similarity ≥ 0.5.

The same matching logic runs retroactively when a new participant joins an event, existing photos are re-scanned against the newcomer’s embeddings.

Storage

Uploaded photos are organized on disk under uploads/<eventId>/. Profile photos (used as reference images) live under uploads/pfps/. The HTTP server serves these files directly, restricting extensions to .jpg, .jpeg, .png, and .webp.

Frontend

Built with SvelteKit and styled with Tailwind CSS. Key routes:

  • /events: list events the user has created or joined
  • /events/[id]: event detail: guest list with upload status, photo grid with face-match overlays
  • /profile: manage reference photos used for face matching
  • /admin: user and event management (admin only)

Face match bounding boxes are rendered as overlays on each photo, tagged with the matched user’s name and confidence score.

Source code