Pretrained Model

[Jitendra Chautharia](IIT Jodhpur)^1,3,

• Python 3.6+
• Nvidia GPU + CUDA, CuDNN

First use the following commands to prepare the environment:

conda create -n ColorVid python=3.6
source activate ColorVid
pip install -r requirements.txt

Then, download the pretrained models from this link, unzip the file and place the files into the corresponding folders:

• video_moredata_l1 under the checkpoints folder
• vgg19_conv.pth and vgg19_gray.pth under the data folder

In order to colorize your own video, it requires to extract the video frames, and provide a reference image as an example.

• Place your Target grayscale image into one folder, e.g., ./exp_sample/target
• Place your reference images into another folder, e.g., ./exp_sample/references

If you want to automatically retrieve color images, you can try the retrieval algorithm from this link which will retrieve similar images from the ImageNet dataset. Or you can try this link on your own image database.

python test.py --image-size [image-size] \
               --clip_path [path-to-target-grayscale-image] \
               --ref_path [path-to-reference] \
               --output_path [path-to-output]

We provide several sample video clips with corresponding references. For example, one can colorize one sample legacy video using:

python test.py --clip_path ./exp_sample/target \
               --ref_path ./exp_sample/references \
               --output_path ./exp_sample/output

Note that we use 216*384 images for training, which has aspect ratio of 1:2. During inference, we scale the input to this size and then rescale the output back to the original size.

We also provide training code for reference. The training can be started by running:

python --data_root [root of video samples] \
       --data_root_imagenet [root of image samples] \
       --gpu_ids [gpu ids] \

We do not provide the full video dataset due to the copyright issue. For image samples, we retrieve semantically similar images from ImageNet using this repository. Still, one can refer to our code to understand the detailed procedure of augmenting the image dataset to mimic the video frames.