facetools

Easy-to-use face related tools, including face detection, landmark localization, alignment & recognition, based on PyTorch.

Quick start

• Do face detection and landmark localization using MTCNN

from PIL import Imagefrom align.detector import detect_facesfrom align.visualization_utils import show_results

img = Image.open('imgs/single.jpg')  # modify the image path to yoursbounding_boxes, landmarks = detect_faces(img)  # detect bboxes and landmarks for all faces in the imageshow_results(img, bounding_boxes, landmarks)  # visualize the results

• Do alignment

from align.face_align import align
res = align('imgs/single.jpg', save_path='./result', vis=False)
res.show()

• Do face encoding using IR50 model (download pretrained model)

from PIL import Imagefrom util.extract_feature import extract_featurefrom backbone.model_irse import IR_50image_1 = Image.open('imgs/align.jpg')  # modify the image path to yoursmodel = IR_50([112, 112])
model_cp = 'checkpoint/backbone_ir50_ms1m_epoch120.pth'features = extract_feature(image_1, model, model_cp)print(features.size())  # output : torch.Size([1, 512])

• Calculate the distance between two images

import numpy as npfrom PIL import Imagefrom util.extract_feature import extract_featurefrom backbone.model_irse import IR_50from scipy.spatial.distance import pdist


face_1 = Image.open('imgs/person_1/17.jpg')
face_2 = Image.open('imgs/person_1/18.jpg')  # face_1 and face_2 belong to the same oneface_3 = Image.open('imgs/person_2/151.jpg')
face_4 = Image.open('imgs/person_2/152.jpg')  # face_3 and face_4 belong to the same onemodel = IR_50([112, 112])
model_cp = 'checkpoint/backbone_ir50_ms1m_epoch120.pth'data = [face_1, face_2, face_3, face_4]

features = extract_feature(data, model, model_cp)
features = [i.numpy() for i in features]  # embeddings for face_1, face_2, face_3 and face_4diff = np.subtract(features[0], features[1])
dist = np.sum(np.square(diff), 1)print(dist)  # output : 1984.6016diff = np.subtract(features[2], features[3])
dist = np.sum(np.square(diff), 1)print(dist)  # output : 1921.2222diff = np.subtract(features[0], features[2])
dist = np.sum(np.square(diff), 1)print(dist)  # output : 16876.32diff = np.subtract(features[1], features[3])
dist = np.sum(np.square(diff), 1)print(dist)  # output : 17107.396dist = pdist(np.vstack([features[0], features[1]]), 'cosine')print(dist)  # output : 0.12932935dist = pdist(np.vstack([features[2], features[3]]), 'cosine')print(dist)  # output : 0.11706942dist = pdist(np.vstack([features[0], features[2]]), 'cosine')print(dist)  # output : 1.09022914dist = pdist(np.vstack([features[1], features[3]]), 'cosine')print(dist)  # output : 1.07447068

• Do face parsing

from PIL import Imagefrom parsing.face_parsing import parsing, vis_parsing_maps

image = Image.open('imgs/9.jpg')

res = parsing(image)
vis_parsing_maps(image, res, show=True, save_im=True)

Using facetools in Your Project

It is easy to use facetools in your project.

Your project
│   README.md
│   ...
│   foo.py
│
└───facetools
│
└───directory1
│   
└───...

In foo.py, you can easily import facetools by adding:

from facetools import detect_faces, show_resultsfrom PIL import Imagedef foo():
    img = Image.open('/path/to/your/image') 
    bounding_boxes, landmarks = detect_faces(img) 
    show_results(img, bounding_boxes, landmarks)

Acknowledgement

• This repo is based on face.evoLVe.PyTorch and face-parsing.PyTorch. Many thanks to the excellent repo.