README.md

Synopsis

The goal of this project is to use Machine Learning to transform police sketches to realistic images.

Motivation

The overall motivation is to help the police better identify and catch the bad guys faster.

Requirements

• imagemagick

Creating Datasets (!Deprecated)

1. Create a folder in sketch2pix/dataset

   • ie. faces-edge1

2. Within that folder create 3 folders edge, face, face2edge

3. Within each of those 3 folders create 2 folders test, train

4. Put your output images in the face folder and split them however you want into the test and train folders. It's recommended you train on 70% of your images and test the other 30%

5. Use the sketch.sh script to generate the edge versions (aka inputs) of your face images

   • in the sketch2pix/dataset run

   ./sketch.sh --image-path /path/to/image_folder --face-path /path/to/face_folder  --sketch-path /path/to/sketch_folder

   • --sketch-path should be the train or test folder in edge

6. Use the combine.sh script to generate the combination images needed for pix2pix to train and test

   • in the sketch2pix/dataset run

   ./combine.sh --path faces-edge1

   • This command will output to the face2edge folder

Creating Dataset

Run

> python sketch2pix/dataset/PencilSketch/gen_sketch_and_gen_resized_face.py \
/path/to/image_folder /path/to/face_folder /path/to/sketch_folder

Training a Model

In sketch2pix

Run

./train.sh --data-root ../dataset/celebfaces/face2edge --name edge2face_generation --direction BtoA --torch /root/torch/install/bin/th

Required parameters:

--data-root: the face2edge folder of your dataset

--name : name of experiment

--direction : BtoA or AtoB

--torch : path to th bin

As the model trains checkpoints will be stored in:

sketch2pix/pix2pix/checkpoints/{--name}

where --name is the value passed into the scripts from above.

More info here

Testing a Model

In sketch2pix

Run

./test.sh --data-root ../dataset/celebfaces/face2edge --name edge2face_generation --direction BtoA --torch /root/torch/install/bin/th [--custom_image_dir my_named_gen]

Required parameters:

--data-root: the face2edge folder of your dataset

--name : should be the same as what was used in the train script

--direction : BtoA or AtoB

--torch : path to th bin

Optional parameters:

--custom_image_dir : name of the directory you want results outputted to. Default uses the model name if not supplied.

More info here

Validating your models

After testing your results are stored in

sketch2pix/pix2pix/results/{--name}/latest_net_G_test/index.html

where --name is the value passed into the scripts from above.

Open that html file in a browser to examine the results of your test.

TODO

On how to improve it

• Realistic sketch generation that matches pencil sketches [done]
• More training data [done]
• Using a face segmenter, seperating the face out of the background, so input photo is just a face on white background. This way, the generated photos will also have a white background and the sketches will not have the background either.
• Using seperate model for men and women, seperate model for different hair colours. All achievable by running through internal FaceSDK

Open Datasets

• Large-scale CelebFaces Attributes
• Labeled Faces In The Wild
• CUHK Face Sketch Database
• CUHK Face Sketch FERET Database
• XM2VTS data set
• SCface - Surveillance Cameras Face Database
• Face Recognition Databases
• adiencedb - homepage

Acknowledgements

• pix2pix
• rakshitw/Sketch

Contributors

• User1m - Script development and Pix2Pix model training / testing, Web App
• koul - Seed idea
• wtam, tikyau, getCloudy - Help with data generation

License

MIT