RetinaFace: Single-stage Dense Face Localisation in the Wild

Face detection in the wild presents unique challenges, but RetinaFace 🌐, a single-stage dense face localization model, tackles them effectively. My latest repository on GitHub explores this model, which enables high-precision facial detection and localization of key landmarks. This implementation integrates several backbone models (MobileNet and ResNet variants), allowing for flexibility between model size and accuracy. 🚀

👉 https://github.com/yakhyo/retinaface-pytorch

RetinaFace is well-suited for real-time applications 🎥 and cross-platform deployment 📲, with support for ONNX exports. The codebase has been refactored for ease of use, and lightweight MobileNet backbones ensure that the model performs efficiently even on low-power devices.

Key Features 🌟

• Lightweight Models: Options like MobileNetV1 (width mult=0.25) for fast inference ⚡
• Improved Accuracy: ResNet backbones deliver higher detection accuracy 🎯
• Webcam & ONNX Support: Real-time inference and flexible deployment 🔄

📈 Performance on WiderFace Dataset

Multi-scale Image Size

RetinaFace Backbones	Pretrained on ImageNet	Easy	Medium	Hard
MobileNetV1 (width mult=0.25)	True	88.48%	87.02%	80.61%
MobileNetV1 (width mult=0.50)	False	89.42%	87.97%	82.40%
MobileNetV1	False	90.59%	89.14%	84.13%
MobileNetV2	True	91.70%	91.03%	86.60%
ResNet18	True	92.50%	91.02%	86.63%
ResNet34	True	94.16%	93.12%	88.90%

Original Image Size

RetinaFace Backbones	Pretrained on ImageNet	Easy	Medium	Hard
MobileNetV1 (width mult=0.25)	True	90.70%	88.12%	73.82%
MobileNetV1 (width mult=0.50)	False	91.56%	89.46%	76.56%
MobileNetV1	False	92.19%	90.41%	79.56%
MobileNetV2	True	94.04%	92.26%	83.59%
ResNet18	True	94.28%	92.69%	82.95%
ResNet34	True	95.07%	93.48%	84.40%

For a detailed setup and to explore various training configurations, check out the https://github.com/yakhyo/retinaface-pytorch and contribute to this project! 👩‍💻👨‍💻