NB: The demo uses YOLOV8-Face model which takes up to 30-40 minutes to compile to TensorRT engine. The first launch takes an enormous time.

The sample demonstrates how to use YOLOV8-Face face detector with landmarks and Adaface face recognition model to build a facial ReID pipeline that can be utilized, for example, in doorbell security systems.

Preview:

The sample is split into two parts: Index Builder and Demo modules.

Tested on platforms:

• Nvidia Turing, Ampere
• Nvidia Jetson Orin family

Index builder module loads images from gallery, detects faces and facial landmarks, performs face preprocessing and facial recognition model inference. The resulting feature vectors are added into hnswlib index, and the index (along with cropped face images from gallery) is saved on disk in the index_files directory.

Note that the Index Builder pipeline processes all images in a single resolution with 16:9 aspect ratio, but the gallery is sent to the pipeline through the Client SDK which can take care of padding the image to a target aspect ratio. As such it's safe to add new images of any aspect ratio to the gallery. The gallery images are expected to contain 1 face each and to be named according to scheme <person_name>_<img_n>.jpeg.

Demo module loads previously generated gallery index file and cropped face images, runs face detection and recognition on a sample video stream, displaying face matches on a padding to the right of the main frame.

git clone https://github.com/insight-platform/Savant.git
cd Savant
git lfs pull
./utils/check-environment-compatible

Note: Ubuntu 22.04 runtime configuration guide helps to configure the runtime to run Savant pipelines.

The demo uses models that are compiled into TensorRT engines the first time the demo is run. This takes time. Optionally, you can prepare the engines before running the demo by using the command:

# you are expected to be in Savant/ directory

./scripts/run_module.py --build-engines samples/face_reid/src/module.yml

Note, there is a bug in the nvv4l2decoder on the Jetson platform so the example currently does not work correctly on that platform. See #314

To build the face reid index, start the index docker compose profile and wait for the services to complete building the index.

# you are expected to be in Savant/ directory

# if x86
docker compose -f samples/face_reid/docker-compose.x86.yml --profile index up

# if Jetson
# currently not supported
docker compose -f samples/face_reid/docker-compose.l4t.yml --profile index up

The first launch can take several minutes as the index-builder-pipeline needs to convert ONNX models into TRT format. Successful module start is indicated by a log messages like

INFO ... > The pipeline is initialized and ready to process data. Initialization took ...
INFO ... > Setting module status to ModuleStatus.RUNNING

After that the index-builder-client container will be started automatically.

The index-builder-client service runs the index_builder_client.py script, which loads the images from the gallery, resizes them to the pipeline frame size while preserving content aspect ratio, sends them to the index-builder-pipeline and uses the received results to create index_files/index.bin index file and the cropped face images in the index_files/processed_gallery dir.

After the services complete, the containers shut down automatically. Check that the index_files directory is created and index.bin file and processed_gallery image directory is written into it.

# you are expected to be in Savant/ directory

# if x86
docker compose -f samples/face_reid/docker-compose.x86.yml --profile demo up

# if Jetson
docker compose -f samples/face_reid/docker-compose.l4t.yml --profile demo up

# open 'rtsp://127.0.0.1:554/stream/video' in your player
# or visit 'http://127.0.0.1:888/stream/video/' (LL-HLS)

# Ctrl+C to stop running the compose bundle

Run the Index Builder according to instructions above.

Download the video file to the data folder. For example:

# you are expected to be in Savant/ directory

mkdir -p data
curl -o data/jumanji_cast.mp4 https://eu-central-1.linodeobjects.com/savant-data/demo/jumanji_cast.mp4

Run the performance benchmark with the following command:

./samples/face_reid/run_perf.sh

Note: Change the value of the DATA_LOCATION variable in the run_perf.sh script if you changed the video.