Swaping the detected face of a given video to a given image.

This repo is based on first order model. The difference is that I added a face detection module for real-time face detection use.

It support python3. To install the dependencies run:

pip install -r requirements.txt conda install -c conda-forge scikit-image

For cropping videos during animation process, you need to install ffmpeg and imageio-ffmpeg:

conda install ffmpeg pip install imageio-ffmpeg

Download vox-adv-cpk.pth.tar from here and put it in ./.

Using PIL to read images and make it as a list.

from PIL import Image img = Image.open(PATH) img_ls = [img1, img2, ...]

from image_animation import FaceDetect detector = FaceDetect() boxes, probs, annotates, faces = detector.detect(img_ls, crop_size=None, mode = 'Extract_largest')

If images in list are not in a same size, set crop_size.

There're 4 modes: Detect_bool, Detect, Extract_largest, Extract_all.

Images of faces will be saved in face_result/faces.

If you want to get the annotation image (face boxes on image), set save_annotate=True and it will be save in face_result/annotations.

Use help(FaceDetect) to see more arguments and details.

from image_animation import Animation animator = Animation() animator.animate(source_video = SOURCE_VIDEO, source_image = 'image_animation/image/einstein.png', result_video = 'result.mp4')

source_video is the video from camera. source_image is the image path of the physicist you want to become, defaults to image_animation/image/einstein.png. And the result will be saved into result_video, defaults to result.mp4.

The run time is about 29.14s with 1 GPU(GTX 1080).