AkashPrasadMishra/VGGFACE-2

VGGFace2 INT8 Face Recognition Model

A high-accuracy INT8 quantized face recognition model optimized for edge devices and challenging image conditions such as low lighting, low resolution, and distorted images.

This model produces 512-dimensional face embeddings and is designed for fast CPU inference using ONNX Runtime.

The model uses a two-stage pipeline:

1. Face Detection & Alignment using buffalo_s from InsightFace
2. Face Recognition using a VGGFace2 INT8 ONNX model

Compared to the standard buffalo_l recognition model, this model shows over 20% higher accuracy in difficult scenarios such as:

• Dim lighting
• Low-quality images
• Compressed images
• Slightly distorted faces
• Long-distance face captures

---

Model Details

Property	Value
Architecture	VGGFace2
Precision	INT8 Quantized
Embedding Size	512
Input Size	160 × 160
Framework	ONNX
Inference Engine	ONNX Runtime
Detection Model	InsightFace buffalo_s
Task	Face Recognition / Verification

---

Key Features

Robust Recognition

Designed to handle:

• dim lighting conditions
• blurred images
• low resolution faces
• partially distorted images

Edge Device Optimization

INT8 quantization provides:

• smaller model size
• faster CPU inference
• lower memory usage
• reduced latency

Suitable for:

• edge AI systems
• embedded devices
• real-time recognition pipelines

---

Face Recognition Pipeline