resolator / facenet-playground

Code for face re-identification task.

facenet-playground

Repo with code for FaceNet (paper) and RetinaFace (paper) usage.

Repo structure

• data - contains trained models (stored with git LFS)
• scripts - contains all useful scripts:
  • common.py - common used functions
  • cut_faces.py - optional script for faces preparation (cutting)
  • demo.py - demonstration script with main functional
  • evaluate.py - script for .tsv files evaluation
  • face_detector.py - helper class to use RetinaFace model
  • face_net.py - helper class to use FaceNet model

Installation

All required python dependencies could be found in requirements.txt

Step-by-step installation in docker + venv

1. Can recommend to use my docker image for further installation:

docker run \
--gpus all \
--shm-size 8G \
-e TERM=$TERM \
-e UNAME=$(whoami) \
-e UID=$(id -u) \
-e GID=$(id -g) \
-e DISPLAY=unix$DISPLAY \
-v /tmp/.X11-unix:/tmp/.X11-unix \
-v /etc/localtime:/etc/localtime:ro \
-v $HOME/:/host_home \
--name facenet-playground \
-it \
resolator/rdi:cuda-10.1-cudnn7-runtime-ubuntu18.04

2. Clone this repo into docker container:

git clone https://github.com/resolator/facenet-playground.git

3. Run install.sh and activate prepared venv:

cd facenet-playground
./install.sh
source .venv/bin/activate

4. Enjoy!

Usage

Use <script_name>.py --help to see scripts usages.

For evaluation by yourself use scripts/evaluate.py. For example:

./scripts/evaluate.py --gt-tsv data/gt.tsv --pd-tsv data/pd.tsv

Evaluation results

• 2020-08-09: Accuracy increased to 76.4%. Achieved with the following command:

./scripts/demo.py --database-dir /data/facenet/datasets/mark/ --testing-dir /data/facenet/datasets/test/ --save-to /data/facenet/datasets/tmp --bs
250 --space l2 --img-size 160 160 --expand-factors 0.25 0.2 0.25 0.05

• 2020-07-28: Current accuracy: 67.88%.