/mtcnn-pytorch-crop

Cropping face region of an image using Multi-task Cascaded Convolutional Networks

Primary LanguageJupyter NotebookMIT LicenseMIT

MTCNN CROP

An MTCNN pytorch implementation to detect and crop face region on an image as described by
Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks.

Instructions

Download the repository and then

crop.py --root-dir path_to_images -output-dir output_path

The system will create an output with the same format as the original path to the images

For examples see test_on_images.ipynb.

Directory Format

The root directory format containing the images of interest need to be held in a similar fashion to:

Aaron_Boyd
  Aaron_Boyd_0001.jpg
  Aaron_Boyd_0002.jpg
  ...

Aaron_Evans
  Aaron_Evans_0001.jpg
  Aaron_Evans_0002.jpg
  ...
  
Aaron_Hornback
  Aaron_Hornback_0001.jpg
  Aaron_Hornback_0002.jpg
  ...
 
 Aaron_Lewis
  Aaron_Lewis_0001.jpg
  Aaron_Lewis_0002.jpg 
  ...

This way, every new identity contains its own sub-directory

Requirements

  • pytorch 0.2
  • Pillow, numpy

Credit

This implementation and algorithm is completely taken from: