Face recognition with Deepstream 6.1

This is a face detection and recognition app build on Deepstream.

Requirements

• Deepstream 6.0+
• GStreamer 1.14.5+
• Cuda 11.4+
• NVIDIA driver 470.63.01+
• TensorRT 8+
• Python 3.6+

Follow deepstream official doc to install dependencies.

Deepstream docker is more recommended.

Pretrained

Please refer to links below for pretrained models and serialized TensorRT engine. Or download from Google driver.

• yolov5
• retinaface
• arcface

Alignment

there should be a face alignment before arcface. Use a custom-gst-nvinfer to preprocess the tensor-meta of retinaface.

Demo custom-gst-nvinfer

Usage

1 - compile retinaface parse function

set a larger face detection bbox so that we can do alignment better later.

cd models/retinaface
make

2 - modify all path in config and codes

• all path in src/config/face/
• line 24 in src/kbds/app/face.py
• video path in test script

3 - Start kafka service before everything.

Then set "ip;port;topic" in codes. e.g.

DSFace("localhost;9092;deepstream")

you can recieve result messages on deepstream topic and error messages on error topic in same ip&port.

4 - start app

cd test
python3 face_test_demo.py

Notice that only support RTSP stream and Gstreamer format.

5 - check result

images before and after alignment will be stored in images.

modify codes in test/cal_sim.py. replace with your own face-feature npy file.

python3 test/cal_sim.py

it will outputs all score.

References

• wang-xinyu/tensorrtx
• NVIDIA-AI-IOT/deepstream_python_apps