Pytorch implementation of ideas from the paper Neural Wireframe Renderer: Learning Wireframe to Image Translations by Yuan Xue, Zihan Zhou, and Xiaolei Huang

• Tested on CentOS 7
• Python >= 3.6
• PyTorch >= 1.0
• TensorboardX >= 1.6

• You can download the data from here. By default, pelease extract all files inside v1.1 to the data/raw_data/imgs folder, and extract all files inside pointlines to the data/raw_data/pointlines folder.
• To preprocess the data, run

python data/preprocess.py --uni_wf

The processed data will be saved under the data folder.

We support both single gpu training and multi-gpu training with Jiayuan Mao's Synchronized Batch Normalization.

Example Single GPU Training

If you are training with color guided rendering:

python train.py --gpu 0 --batch_size 14

If you are training without color guided rendering:

python train.py --gpu 0 --batch_size 14 --nocolor

Example Multiple GPU Training

python train.py --gpu 0,1,2,3 --batch_size 40

Tensorboard Visualization

tensorboard --logdir results/tb_logs/wfrenderer --port 6666

Note that the --nocolor option needs to be used consistently with training. For instance, you cannot train with --nocolor and test without --nocolor.

python test.py --gpu 0 --model_path YOUR_SAVED_MODEL_PATH --out_path YOUR_OUTPUT_PATH

For now we only support rasterized wireframes as input, we will release the vectorized wireframe version in the near future.

We hope our implementation can serve as a baseline for wireframe rendering. If you find our work useful in your research, please consider citing:

@inproceedings{xue2020neural,
  title={Neural Wireframe Renderer: Learning Wireframe to Image Translations},
  author={Xue, Yuan and Zhou, Zihan and Huang, Xiaolei},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2020}
}

Part of our code is adapted from CycleGAN. We also thank these great repos utilized in our code: LPIPS, MSSSIM, SyncBN,