RojakFace 1.0.0

A modern face recognition library with automatic learning, vector search, and AI assistant support

RojakFace

A modern face recognition library with automatic learning, vector search, and AI assistant support.

Built with YuNet (deep-learning face detector) and CLIP-based embeddings (imgbeddings) for accurate face detection and recognition.

Features

• YuNet Face Detector — Deep-learning ONNX model from OpenCV Zoo. More accurate than Haar cascades, detects faces at various angles and sizes.
• Automatic Learning — Upload a group photo and it will detect every face, recognize known identities, and automatically learn new ones with unique IDs.
• Smart Deduplication — Two-phase batch processing prevents cross-matching within the same image. Each face is compared against the existing database before any new vectors are inserted.
• Hybrid Storage — PostgreSQL for metadata and sighting records, Qdrant for 768-d vector similarity search (cosine distance).
• Pluggable Storage — Swap between local filesystem and MinIO (S3-compatible) without changing code.
• AI-Ready — Built for LLM integration. Provides structured identity context with image URLs for AI assistants (Ollama, OpenAI, etc.).
• Identity Renaming — Rename across both PostgreSQL and Qdrant with one call.

Installation

pip install RojakFace

Quick Start

import RojakFace as RF
import psycopg2
from qdrant_client import QdrantClient

# 1. Setup connections
pg_conn = psycopg2.connect(database="faces_db", user="dev", password="password")
q_client = QdrantClient("localhost", port=6333)

# 2. Initialize manager
manager = RF.RojakFaceManager(pg_conn, q_client)
manager.init_storage()

# 3. Upload and auto-learn
stats = manager.auto_learn("group_photo.jpg", filename="family.jpg")
print(f"{stats['new_faces']} new faces, {stats['known_faces']} known")

# 4. Rename an identity
manager.update_identity("Person_abc123", "Mom")

# 5. Recognize without saving
results = manager.recognize("new_photo.jpg")
for face in results:
    print(f"{face['identity']} ({face['confidence']:.2f})")

Storage

LocalStorage (default)

storage = RF.LocalStorage(base_path="/path/to/images")
manager = RF.RojakFaceManager(pg_conn, q_client, storage=storage)

MinIO (S3-compatible)

storage = RF.MinioStorage(
    endpoint="localhost:9000",
    access_key="minioadmin",
    secret_key="minioadmin",
    bucket_name="rojakface",
)
manager = RF.RojakFaceManager(pg_conn, q_client, storage=storage)

Custom Storage

from RojakFace import BaseStorage

class MyStorage(BaseStorage):
    def save_image(self, image_data, folder, filename): ...
    def get_url(self, folder, filename): ...

API

RojakFace

Low-level detector with YuNet + imgbeddings.

detector = RF.RojakFace()
results = detector.extract_features("image.jpg")
# Returns: [{ 'box': [x,y,w,h], 'confidence': 0.92, 'embedding': [...], 'face_image': np.array }]

RojakFaceManager

Method	Description
init_storage()	Create PostgreSQL table and Qdrant collection
auto_learn(image, threshold=0.85, filename=None)	Detect, recognize known, learn new
recognize(image, limit=1, threshold=0.7)	Recognize without saving
update_identity(old, new)	Rename across all storage backends

Requirements

• Python >= 3.8
• PostgreSQL (with psycopg2)
• Qdrant vector database
• Ollama (optional, for AI chat features)

---

Made by lqmnwido