deepstream-face-recognition

Face detection -> alignment -> feature extraction with deepstream

1. Requirements

• Deepstream 6.1 (recommend using docker)

2. Installation

2.1 Build docker

docker build -t face_recog_ds6.1 - < Dockerfile

2.2 Run docker (remember to mount)

docker run --rm -it -v<path-to-source>:<path-to-destination> --gpus all face_recog_ds6.1:latest

2.3 Compile plugins (run inside docker)

• Compile custom NMS plugin (batchedNMSDynamic_TRT with landmark)

  export CUDA_VER=11.4
  cd plugins/nms
  mkdir build && cd build
  cmake ..
  make

• Compile custom gst-nvinfer plugin

  cd gst-nvinfer
  make
  make install

  3. Run demo

• object detection -> face detection (from cropped person image) -> face embedding

  LD_PRELOAD=<path-to-NMS-plugin> python main.py file:<path-to-video-input>

• face detection (full frame) -> face embedding

  LD_PRELOAD=<path-to-NMS-plugin> python main_ff.py file:<path-to-video-input>

4. To do

• [x] Add cropped/fullframe pipeline for face
• [x] Fix 1-batchsize error (we have some bugs with custom NMS plugin, only have true output with batchsize=1)
• [ ] Add GPU face alignment
• [ ] Add cropped pipeline for license plate