release v0.0.1

In this implementation, we use several lightweight and powerful backbone architectures to provide flexibility between performance and accuracy.

📈 Results on WiderFace Eval

Results of RetinaFace (MXNet-based Image Size)

RetinaFace Backbones	Pretrained on ImageNet	Easy	Medium	Hard	#Params (M)
MobileNetV1 (width mult=0.25)	False	88.48%	87.02%	80.61%	0.42
MobileNetV1 (width mult=0.50)	False	89.42%	87.97%	82.40%	1.65
MobileNetV1	False	90.59%	89.14%	84.13%	4.16
MobileNetV2	True	91.70%	91.03%	86.60%	3.12
ResNet18	True	92.50%	91.02%	86.63%	12.01
ResNet34	True	94.16%	93.12%	88.90%	22.12
ResNet50	True				27.29

Results of RetinaFace, based on Original Image Size

RetinaFace Backbones	Pretrained on ImageNet	Easy	Medium	Hard	#Params (M)
MobileNetV1 (width mult=0.25)	False	90.70%	88.12%	73.82%	0.42
MobileNetV1 (width mult=0.50)	False	91.56%	89.46%	76.56%	1.65
MobileNetV1	False	92.19%	90.41%	79.56%	4.16
MobileNetV2	True	94.04%	92.26%	83.59%	3.12
ResNet18	True	94.28%	92.69%	82.95%	12.01
ResNet34	True	95.07%	93.48%	84.40%	22.12
ResNet50	True				27.29

✨ Features

• ✅ Cleaner & Reproducible Code: Refactored for simplicity and consistency, making it easier to use and maintain.
• 📱 MobileNetV1_0.25 & MobileNetV1_0.50: Lightweight versions for faster inference with reduced computational cost.
• 📲 MobileNetV1: Efficient Convolutional Neural Networks for Mobile Vision Applications - Optimized for mobile and low-power applications.
• 📈 MobileNetV2: Inverted Residuals and Linear Bottlenecks - Improved efficiency for mobile use-cases with advanced architecture.
• 🔍 ResNet Models (18, 34, 50): Deep Residual Networks - Enhanced accuracy with deeper residual connections, supporting a range of model complexities.