Final Project

Lucas Tindall, ltindall@ucsd.edu

Samuel Sunarjo, ssunarjo@eng.ucsd.edu

Abstract Proposal

In this project we build a system for real time translation of human faces to cartoon stylized faces. We trained CycleGANs on different cartoon styles to perform the transformations. The end product is a real time application which displays a camera feed where real human faces are replaced by the selected cartoon style.

Project Report

Report Link

Model/Data

• CycleGAN - Folder containing the Pytorch implementation of CycleGAN written by junyanz
• CartoonGAN - Folder containing the Pytorch implementation of CartoonGAN written by znxlwm
• model checkpoints - Model files used in the generation scripts below.

The datasets we used for training can be found in the following repositories

• CaVI Dataset: includes the real human faces used in all models and the hand drawn caricatures
• Anime Face Dataset: includes anime faces
• Simpsons Faces Dataset: includes faces from The Simpsons
• Cartoon Set Dataset: includes cartoon avatar faces

Code

• camera_gen.py - Script to overlay transformations on a web cam video feed.

cd pytorch-CycleGAN-and-pix2pix
python camera_gen.py --dataroot . --load_size 129 --crop_size 128 --no_dropout --gpu_ids -1

• generate.py - Script to generate transformations on a static image.

cd pytorch-CycleGAN-and-pix2pix
python generate.py  --dataroot . --load_size 140 --crop_size 128 --no_dropout --gpu_ids -1

Results

WebCam video - An example video(.avi) of the webcam feed overlayed with the cartoon style transformations.

Technical Notes

Runs on the datahub.

Reference

Pytorch implementation of CycleGAN

Pytorch implementation of CartoonGAN

Image-to-Image Translation with Conditional Adversarial Networks