Using CycleGAN with building sketches + images
To train a CycleGAN model, you need two sets of images representing the two archetypes you want to translate between (in our case, images of buildings and sketches of buildings). For convenience, you can place them in the datasets folder under trainA
and trainB
.
For best results, the images should be the same size (we used 512x512px). You can use dataset-tools to quickly resize, crop, or stretch images to a common size and format.
Using a NVIDIA GPU with CUDA Deep Neual Network (cuDNN) installed (we used a machine from Paperspace):
Clone the pyTorch implementation of CycleGAN:
git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
cd pytorch-CycleGAN-and-pix2px
Install depedencies:
pip install -r requirements.txt
Train the model:
python train.py --dataroot /path/to/datasets --name sketch2building --model cycle_gan
NOTE: This process can takes 24-48 hours, depending on the specs of the machine + GPU you are using to train it.
TBD: Still figuring this one out 😅
Update: Used Pix2PixHD model (a much more advanced CoGAN than Pix2Pix which uses semantic segmentation in the input labels to drastically improve the results). It was not possible to do the same for our custom dataset, but there were dramatic improvements by increasing the size of training dataset and restricting the images to mostly one or two type of objects. Some of the test results:
OpenCV for dataset processing from images
CycleGAN for training
dataset-tools by dvschultz
Thanks to Derrick Schultz for sharing his lectures on the CycleGAN process. Check out his work here.