Face counting

Pytorch0.4.1 codes for counting the number of unique faces under a directory containing images.

1. Intro

• This repo is heavily borrowed from Insightface_Pytorch

• This code also save the identity embeddings of the unique faces found as an .npz file, where the keys are

  embeddings: the identity embeddings, PyTorch tensor of shape (num_embeddings, embed_dim)  
  names: the names of the embeddings, array of shape (num_embeddings + 1,) where the first being "Unknown"
  emb_counts: number of occurances of each embeddings, array of shape (num_embeddings,)
  

• Pretrained weights. (Note: I used IR-SE50) For InsightFace IR-SE50 @ Onedrive For RetinaFace RetinaFace-R50

2. How to use

2.1. Pre-requisites

• clone this repo

  git clone https://github.com/yongxinw/face_counting.git
  

• install requirements by (requires python 3.6)

  pip install -r requirements.txt
  

• set environmental variables

  export LD_LIBRARY_PATH=<path_to_cuda_installation>/lib64:$LD_LIBRARY_PATH
  export LD_LIBRARY_PATH=<path_to_cuda_installation>:$LD_LIBRARY_PATH
  

• download pretrained weights by clicking the links from section 1

2.2. Run

2.2.1. RetinaFace for detection

• go to the RetinaFace directory, and compile RetinaFace

  cd <path_to_this_repo>/RetinaFace
  make
  

• run RetinaFace detection

  python test_folder.py \
  --image_root <directory containing images> \
  --res_root <directory to save detection visualizations> \
  --model_path <path to the pretrained model>
  

2.2.2. InsightFace for verification

• run the following command to count the number of faces under a directory

  cd <path_to_this_repo>
  python count_faces.py \
  --result_root <directory to save the .npz file> \
  --retinaface <path to .txt file containing the RetinaFace detection> \ 
  --image_root <directory of the images> \
  --model_path <directory containing the insightface pretrained model> \
  --verbose
  

3. References

• This repo is heavily borrowed from Insightface_Pytorch
• See also deepinsight/insightface