kodamap / face_reidentification_demo

• Face Re-identification with OpenVINO
  • What's this
  • Reference
  • Environment
  • Required Python packages
  • How to use
  • Create Face Database
    • 1. Download Face images from CelebA
    • 2. Create CSV file
    • 3. Regiter command
    • 4. Verify data
  • Run app

Face Re-identification with OpenVINO

What's this

This is Face Identification Test App with Intel OpenVINO Face Re-Identification Model.

You can do followings:

• Realtime Face Re-Identification
• Face Search

Real time face re-identifiction (Youtube Link)

Face Search

Reference

• Install OpenVINO Toolkit

• face-reidentification-retail-0095

• CelebA Dataset

• Flask Video streaming

Environment

• Python 3.6+ (Required Ordered Dict)
• OpenVINO Toolkit 2021.4[^1]
• Windows 10

Required Python packages

pip install -r requirements.txt

How to use

python app.py -h

usage: app.py [-h] -i INPUT [-l CPU_EXTENSION] [-d {CPU,GPU,FPGA,MYRIAD}]
              [-d_lm {CPU,GPU,FPGA,MYRIAD}] [-d_fi {CPU,GPU,FPGA,MYRIAD}]
              [--dbname DBNAME] [--no_v4l]

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Path to video file or image. 'cam' for capturing video
                        stream from camera
  -l CPU_EXTENSION, --cpu_extension CPU_EXTENSION
                        MKLDNN (CPU)-targeted custom layers.Absolute path to a
                        shared library with the kernels impl.
  -d {CPU,GPU,FPGA,MYRIAD}, --device {CPU,GPU,FPGA,MYRIAD}
                        Specify the target device for Face Detection to infer
                        on; CPU, GPU, FPGA or MYRIAD is acceptable.
  -d_lm {CPU,GPU,FPGA,MYRIAD}, --device_landmarks {CPU,GPU,FPGA,MYRIAD}
                        Specify the target device for Facial Landmarks
                        Estimation to infer on; CPU, GPU, FPGA or MYRIAD is
                        acceptable.
  -d_fi {CPU,GPU,FPGA,MYRIAD}, --device_reidentification {CPU,GPU,FPGA,MYRIAD}
                        Specify the target device for Facial re-identificaiton
                        to infer on; CPU, GPU, FPGA or MYRIAD is acceptable.
  --dbname DBNAME       Specify face database name
  --no_v4l              cv2.VideoCapture without cv2.CAP_V4L

Create Face Database

At first, you need to create face Database.

Here is an example of face registration from csv file.

1. Download Face images from CelebA

Download dataset from CelebA using Google Drive [CelebA > img > img_align_celba.zip] and extract the file.

Note: img_align_celba.zip includes 202,599 face images and 1GB size

2. Create CSV file

I tested to register 20,000 faces. Change /path/to/celeba.

• celeba.csv

imagepath, label
<path_to_celeba_img_align_celeba>\000001.jpg,F00001
<path_to_celeba_img_align_celeba>\000002.jpg,F00002
<path_to_celeba_img_align_celeba>\000003.jpg,F00003
..
<path_to_celeba_img_align_celeba>\000004.jpg,F00004

3. Regiter command

[dbname]_vecs.gz and [dbname]_pics.gz are created. (You can ignore the errors during face registration.)

python registrar.py csv_register --csv celeba.csv --dbname celeba --batch_size 500

celeba_vecs.gz includes feature vectors, celeca_pics.gz includes image path of each face.

The size of feature vectors file produced from 20,000 face images is about 22 MB.

>dir
..
2019/06/26  22:06           159,470 celeba_pics.gz
2019/06/26  22:06        23,168,652 celeba_vecs.gz

Face images are saved to /static/images/dbname

4. Verify data

python registrar.py list --dbname celeba
..
19781, label:F19999 file:/static/images/celeba/F19999.png
19782, label:F20000 file:/static/images/celeba/F20000.png
Rows:19782

Run app

#  dbname is the face db name (ex. celeba) 
python app.py -i cam --no_v4l --dbname celeba

If you use OpenVINO Toolkit 2019 R3 or earlier build please specify cpu_extension.

#  dbname is the face db name (ex. celeba) 
python app.py -i cam -l extension\cpu_extension.dll --no_v4l --dbname celeba

Access to the url

http://127.0.0.1:5000/