🤖 RetinaFace Pro

Face Detection · Landmark Localisation · Identity Verification

Industry-grade computer vision for dense face detection, precise 5-point landmark localisation, and one-to-many identity verification — production-ready and deployed live.

🚀 Try Live Demo · 📦 HuggingFace Space · 👤 Author

---

📖 Overview

RetinaFace Pro is a high-performance computer vision pipeline designed for dense face detection and identity verification. Unlike simple research scripts, this project is built as a modular ML system wrapping state-of-the-art models into a professional engineering architecture.

🚀 The Engineering Value-Add

• ✅ Production Wrapper: Clean, documented Python API (src/detector.py).
• ✅ ML Monitoring: Integrated MLflow tracking for inference latency and confidence drift.
• ✅ Reliability: 100% testable logic with Pytest and automated CI/CD.
• ✅ Portability: Fully Dockerized for zero-friction deployment on Hugging Face Spaces.

---

📊 Technical Benchmarks

Tested on: NVIDIA GTX 1650 GPU | 16GB RAM | Intel i7

Task	Latency (ms)	Throughput (FPS)
Single Face Detection	~45ms	~22 FPS
Multi-Face (5+) Detection	~65ms	~15 FPS
Identity Verification	~110ms	~9 FPS

---

🏗️ Architecture Decisions

To build a "Top Class" system, we prioritized Engineering Trade-offs over blind model selection. Here is the Why behind our choices:

1 · RetinaFace over Two-Stage Detectors

While two-stage detectors (Faster R-CNN) are accurate, they are often too slow for real-time inference. We chose RetinaFace because its Feature Pyramid Network (FPN) handles multi-scale faces (tiny to large) in a single pass, eliminating the need for a redundant Region Proposal Network (RPN) overhead.

2 · ArcFace for Verification Accuracy

Standard Softmax loss functions struggle with face verification because they don't optimize for embedding compactness. We implemented ArcFace (Additive Angular Margin Loss) because it enforces a tighter distance between similar faces in the hyperspace, leading to superior identity separability compared to standard cosine similarity.

3 · MobileNet-0.25 (The Latency-Accuracy Tradeoff)

To achieve our ~45ms inference latency on a GTX 1650, we used the MobileNet-0.25 backbone. It offers the best performance-per-watt, ensuring the system remains responsive even in dense multi-face scenarios without requiring expensive A100 GPUs.

4 · The Necessity of 5-Point Alignment

Verification accuracy drops by ~15-20% if faces are not aligned. Our pipeline performs an explicit 5-point landmark transformation (eyes, nose, mouth corners) before generating embeddings, ensuring that the ArcFace backbone receives spatially consistent input.

---

🏗️ Project Structure

Ratina_Face/
├── src/                    # Core logic — FaceDetector wrapper
├── tests/                  # Automated test suite (pytest)
├── .github/workflows/      # CI/CD pipelines (GitHub Actions)
├── app.py                  # Streamlit web application
├── main.py                 # Command-line tool
├── Dockerfile              # Production container setup
└── requirements.txt        # Managed dependencies

---

📦 Quick Start

1 · Clone & Install

git clone https://github.com/ImdataScientistSachin/RetinaFace-Detection
cd RetinaFace-Detection
python -m pip install -r requirements.txt

2 · Operational Monitoring (MLflow)

To view inference logs and latency benchmarks:

mlflow ui

3 · Deployment

docker build -t retinaface-pro .
docker run -p 7860:7860 retinaface-pro

---

👤 Author

Sachin Paunikar — LinkedIn | GitHub

_{Built with ❤️ · Powered by RetinaFace & DeepFace · Deployed on HuggingFace Spaces}