arcface-embedder

Simple script for face detection, alignment and getting embeddings using pretrain models from insightface project, converted to PyTorch format using pytorch-insightface project and using detector from MTCNN project. Face alignment implemented using pytorch Tensor computing, based on original insightface numpy realization

Install/build dependencies (FaceDetector and pytorch-insightface)

1. Load submodules:

$ git submodule init

or just use one-step method during clonnig project:

$ git clone --recursive https://github.com/MZHI/arcface-embedder.git

2. Build FaceDetector submodule using instructions from FaceDetector/README.md:

$ cd FaceDetector/
$ pip install opencv-python numpy easydict Cython progressbar2 torch tensorboardX
$ python setup.py build_ext --inplace
$ python setup.py install

3. Install insightface module:

$ pip install git+https://github.com/nizhib/pytorch-insightface

4. If you want to get weights for embedder locally, you need to convert weights from model zoo, 3.1, 3.2 and 3.3, follow instructions from pytorch-insightface/README.md: download the original insightface zoo weights and place *.params and *.json files to pytorch-insightface/resource/{model}. Then run python pytorch-insightface/scripts/convert.py to convert and test pytorch weights.

Usage

You can just run script without any parameters, all available input parameters will be setted to default values, see next section for details:

$ python3 run.py

For running with specific image path:

$ python3 run.py --image-path [image name]

Options

There are some input parameters available:

• --image-path: path to image to be processed. Default: ./images/office5.jpg
• --is-local-weights: whether to use local weights or from remote server. Default: 0
• --weights-base-path: root path to insightface weights, converted to PyTorch format. Actual only if --is-local-weights == 1. Default: pytorch-insightface/resource
• --show-face: whether to show cropped and aligned face or not. Default: 0
• --align-torch: whether to use torch or numpy realization for face alignment. Default: 1
• --arch: architecture of embedder: iresnet34|iresnet50|iresnet100. Default: iresnet100

Some examples of aligning and distance between features

For comparing two vectors of features L1 norm was used:

numpy.linalg.norm(v1-v2, 1)

where v1 and v2 are feature vectors of size 512

detected original face	aligned image 1	aligned image 2	L1-norm
	torch align:	numpy align:	2.4723
	torch align:	numpy align:	2.3088
	torch align:	numpy align:	1.5965
---	torch align:	torch align:	29.0398
---	torch align:	torch align:	23.1007
---	torch align:	torch align:	27.0426