Automatic Mafia

A desktop application that registers participants of a mafia game from a local camera, keeps a stable player number and face identity for each of them, and gives the host a control flow for the game. A second window works as a public board.

The vision part runs in worker processes and never writes game actions. The game state lives in a single store, and the public board only receives filtered public data.

GitHub repository

Category :Desktop App / Computer Vision
Stack :Python, PySide6, OpenCV, face_recognition, dlib
Windows :Host control window + public board
Identity :Stable numbers + face database
Data :Local files only
Source :GitHub repository
Status :Prototype complete

Automatic Mafia table with detected players

Core Features

The prototype turns a physical party game into a managed session: the camera handles who is who, the host handles the game.

Registration of participants from a local camera.
Stable player numbers and face identities for the whole session.
Position-based track holding when a face is briefly lost.
A manual host flow for roles, phases and night actions.
A separate public board that shows only public data.
Elimination reasons stored per participant.

“Vision must not write game actions: the game store is the only source of truth.”

Design Note - Separated Responsibilities

Architecture

Two Qt windows run in one process. A camera worker handles capture and face registration with OpenCV and face_recognition, and feeds the session store with participants, photos and stable numbers. A game store owns roles, phases and night rules. A controller connects them to the host window and to the public board, filtering what the board may show. All data stays in local files.