Edge / Jetson Deployment

Edge Deployment

Run OpenBiometrics on NVIDIA Jetson (Nano, Orin) and other embedded devices.

Jetson (Docker)

docker build -f Dockerfile.jetson -t openbiometrics:jetson .
docker run -p 8000:8000 --runtime nvidia openbiometrics:jetson

The Jetson image uses TensorRT for optimized inference, achieving 2-5x speedup over CPU.

Supported Platforms

Platform	Runtime	Notes
Jetson Nano/Orin	TensorRT	Best performance on NVIDIA edge
Raspberry Pi 5	ONNX Runtime	CPU only, suitable for low-throughput
Intel NUC	OpenVINO	Optimized for Intel CPUs
Generic ARM	NCNN	Minimal footprint (~700KB)

Model Optimization

For edge deployment, consider:

1. Quantize to INT8 — 4x smaller models, minimal accuracy loss
2. Use smaller backbones — MobileFaceNet instead of ResNet-50
3. Reduce detection resolution — 320x320 instead of 640x640

# Convert ONNX to TensorRT (on Jetson)
import tensorrt as trt
# See engine/scripts/convert_tensorrt.py