RetinaFace – Accurate AI Face Detection & Facial Analysis

91.4%

AP on WIDER FACE Hard Subset

100+

FPS on GPU with MobileNet Backbone

5

Facial Landmarks Detected per Face

1

Single Forward Pass — No Multi-Stage Pipeline

🎯

Single-Stage Architecture

Detect faces in one forward pass with no multi-stage cascade, delivering fast inference without sacrificing accuracy.

📐

Five-Point Landmark Detection

Outputs precise coordinates for both eye centers, nose tip, and both mouth corners for accurate face alignment.

🔭

Multi-Scale Detection

Feature Pyramid Network enables detection of faces from 16×16 pixels to full-frame close-ups in a single inference.

🧠

Multi-Task Learning

Simultaneously trained on face classification, bounding box regression, and landmark localization for richer representations.

🌐

3D Mesh Reconstruction

Optional dense 3D face mesh branch provides head pose, depth, and structural face data for AR and animation.

📱

Lightweight Backbone Options

Choose MobileNet for CPU and edge deployment or ResNet-50 for maximum accuracy on GPU servers.

⚡

Real-Time Performance

Achieves 100+ FPS on modern GPUs with the MobileNet backbone — suitable for live video and streaming.

🔄

ONNX & Framework Portable

Export to ONNX and run on TensorRT, OpenVINO, or ONNX Runtime for maximum deployment flexibility.

📊

Benchmark-Leading Accuracy

Achieves 91.4% AP on the WIDER FACE hard subset, outperforming MTCNN, PyramidBox, and many two-stage detectors.

1

⬇

Download the Package

Click the download button to get the RetinaFace model package including pre-trained weights and Python scripts.

2

🐍

Install Python Dependencies

Run pip install retinaface in your terminal. Dependencies including NumPy and OpenCV install automatically.

3

📂

Extract the Files

Unzip the downloaded package into your project directory. The folder structure is pre-organized for immediate use.

4

⚙

Load the Model

Import RetinaFace in your Python script and call the detect_faces method with your image path or NumPy array.

5

🖼

Run Detection

Pass any image to RetinaFace and receive bounding boxes, confidence scores, and five-point landmark coordinates for every detected face.

6

🔗

Integrate Your Pipeline

Feed the detection output into your downstream model — face recognition, alignment, analysis, or any computer vision task.

🤖

RetinaFace Model Package

Pre-trained weights, Python implementation, and usage examples in one clean download.

⬇ Download Now — Free

✓ Free Forever ✓ No Sign-up ✓ MIT License

👨‍💻

ML Engineers

Integrate a production-ready face detection layer into your recognition or analysis pipeline without reinventing the wheel.

🎓

Researchers

Use RetinaFace as a reliable baseline or benchmark starting point for computer vision and face analysis research.

📸

App Developers

Add automatic face detection and tagging to photo apps, social platforms, and media management tools.

🔐

Security Teams

Deploy RetinaFace in access control, surveillance, and identity verification systems requiring reliable multi-face detection.

🎮

AR/VR Developers

Use the 3D mesh output to drive avatar animation, virtual try-on, and AR filter applications accurately.

🏥

Healthcare AI

Apply RetinaFace in dermatology screening, ophthalmology tools, and patient monitoring systems needing precise facial localization.

★★★★★

"RetinaFace replaced MTCNN in our entire production pipeline and improved our end-to-end recognition accuracy by over 4%. The landmark predictions are noticeably more stable on challenging angles."

★★★★★

"We deployed RetinaFace with the MobileNet backbone on Jetson Nano devices across 12 retail locations. It handles real crowds reliably and has been running without issues for six months."

★★★★★

"The ONNX export workflow is smooth and well-documented. Moving from PyTorch training to TensorRT inference cut our detection latency from 18ms to 6ms per frame."

★★★★

"Excellent detection on the WIDER FACE hard set cases. I tried several detectors for my research on crowd analysis and RetinaFace gave the best recall on densely packed scenes by a clear margin."

★★★★★

"Integration took less than an hour. The Python API is clean, the docs are clear, and the performance exceeded what I expected from an open-source model. Highly recommended for any face pipeline."

★★★★★

"The five-point landmarks are accurately calibrated even on faces with mild occlusion. Our face alignment quality improved significantly after switching from a two-stage detector to RetinaFace."