This repository provides the code for the WACV20 paper Synthesizing human-like sketches from natural images using a conditional convolutional decoder. Find more details Here.

Setting up the environment

The repo runs with the included requirements.txt.

Pretrained Models

Download pretrained models from here and put them in folder pretrained_models

Testing on a single image

After downloading the pretrained models, simply run python test.py --img_path <path-to-image> --label <label-of-image>. See python test.py --help for the list of parameters.

Training

First download the data and make sure you have the pretrained models resnet_classifier.pt and PSim_alexnet.pt downloaded inside pretrained_models.

Data preparation

Download the sketchy database from http://sketchy.eye.gatech.edu/. Get both the Sketches and Photos as well as the Annotation and Info.

Unpack into the following structure:

+-- data/
|   +-- photo/
|   +-- sketch/
|   +-- info/

Run the training

For training your own network simply run python train.py, it runs with default parameters. See python train.py --help for the list of parameters. The training script outputs to stdout and generates logfiles as well as tensorboard output.

Citation

If you use anything from this work for your own, please cite

@inproceedings{kampelmuehler2020synthesizing,
  title={Synthesizing human-like sketches from natural images using a conditional convolutional decoder},
  author={Kampelm{\"u}hler, Moritz and Pinz, Axel},
  booktitle={IEEE Winter Conference on Applications of Computer Vision (WACV)},
  year={2020}
}