harry-thasarathan / tcvc Goto Github PK

ArXiv

We develop a new approach to automatically colorize line art/greyscale anime episode frames to a temporally coherent video using an end-to-end GAN based learning framework. In the animation industry, heavy workloads and overtime are the norm for young artists. When between 3000 to 10000 frames need to be individually animated for a single episode, the work needs to be distributed and outsourced which leads to slow turn around. This method is our initial foray at producing an algorithm that could benefit animation studios and reduce workloads. We mimic lineart using Canny edge detection and adapt a training scheme to preserve temporal information between generated frames to mitigate the flicker effect that exists when using GANs to generate individual frames for video. Detailed information of our work can be found in our paper.

(Left) Ground truth colored frames. (Middle) Computed sythetic linearts. Edges drawn in black are computed using Canny edge detector. (Right) Colored frame prediction produced by generator.

• Python 3
• PyTorch 1.0
• NVIDIA GPU + CUDA cuDNN

• Clone this repo:

git clone https://github.com/harry-thasarathan/TCVC.git
cd TCVC

• Install PyTorch and dependencies from http://pytorch.org
• Install python requirements:

pip install -r requirements.txt

In order to train our automatic video colorization network for anime, we generate a dataset of 60k frames from about 8 episodes of season 1 of the original Dragonball TV show obtained from legal sources. To do so we wrote a script FrameExtraction.py in which a path to the video file and output folder for the extracted frames can be specified. You can use the same script to extract frames for your own favorite tv show episodes by specifying the video file paths.

python FrameExtraction.py --video_path [video path] --output_path [output path]

To train the model, place the dataset folder in the same folder of your repository which will be used as the root path during training and testing. Run the following in terminal to begin the training process. Checkpoints are saved every epoch by default and samples are produced in the root directory every 250 iterations. Refer to the argument parser in Train.py for more configurability options.

python train.py \
  --root [path to root directory (location of repo)] \
  --dataset [name of dataset folder in root directory] \
  --logfile [name of logfile to be generated to keep track of losses] \
  --checkpoint_path_G [loading a pretrained generator] \
  --checkpoint_path_D [loading a pretrained discriminator] \
  --batchSize [size of batch] \

If you use this code for your research, please cite our paper TCVC: Temporally Coherent Video Colorization: