Skip to content

CharacterGAN: Few-Shot Keypoint Character Animation and Reposing (Best Paper WACV 2022)

License

Notifications You must be signed in to change notification settings

tohinz/CharacterGAN

Repository files navigation

CharacterGAN

Implementation of the paper "CharacterGAN: Few-Shot Keypoint Character Animation and Reposing" by Tobias Hinz, Matthew Fisher, Oliver Wang, Eli Shechtman, and Stefan Wermter (open with Adobe Acrobat or similar to see visualizations).

Supplementary material can be found here.

Our model can be trained on only a few images (e.g. 10) of a given character labeled with user-chosen keypoints. The resulting model can be used to animate the character on which it was trained by interpolating between its poses specified by their keypoints. We can also repose characters by simply moving the keypoints into the desired positions. To train the model all we need are few images depicting the character in diverse poses from the same viewpoint, keypoints, a file that describes how the keypoints are connected (the characters skeleton) and which keypoints lie in the same layer.

Examples

Animation: For all examples the model was trained on 8-15 images (see first row) of the given character.

Training Images 12 15 9 12 15 8 8
Animation dog_animation maddy_animation ostrich_animation man_animation robot_animation man_animation cow_animation



Frame interpolation: Example of interpolations between two poses with the start and end keypoints highlighted.

man man man man man man man man man man man man man
dog dog dog dog dog dog dog dog dog dog dog dog dog



Reposing: You can use our interactive GUI to easily repose a given character based on keypoints.

Interactive dog_gui man_gui
Gui cow_gui man_gui

Installation

  • python 3.8
  • pytorch 1.7.1
pip install -r requirements.txt

Training

Training Data (See datasets Folder for Examples)

All training data for a given character should be in a single folder. We used this website to label our images but there are of course other possibilities.

The folder should contain:

  • all training images (all in the same resolution),
  • a file called keypoints.csv (containing the keypoints for each image),
  • a file called keypoints_skeleton.csv (containing skeleton information, i.e. how keypoints are connected with each other), and
  • a file called keypoints_layers.csv (containing the information about which layer each keypoint resides in).

The structure of the keypoints.csv file is (no header): keypoint_label,x_coord,y_coord,file_name. The first column describes the keypoint label (e.g. head), the next two columns give the location of the keypoint, and the final column states which training image this keypoint belongs to.

The structure of the keypoints_skeleton.csv file is (no header): keypoint,connected_keypoint,connected_keypoint,.... The first column describes which keypoint we are describing in this line, the following columns describe which keypoints are connected to that keypoint (e.g. elbow, shoulder, hand would state that the elbow keypoint should be connected to the shoulder keypoint and the hand keypoint).

The structure of the keypoints_layers.csv file is (no header): keypoint,layer. "Keypoint" is the keypoint label (same as used in the previous two files) and "layer" is an integer value desribing which layer the keypoint resides in.

See our example training data in datasets for examples of both files.

We provide two examples (produced by Zuzana Studená) for training, located in datasets. We also provide the Dino example (taken from https://www.gameart2d.com) from which we used the 10 "Walk" images for training and the 8 training images for the "Fox" example. Our other examples were trained on data from Adobe Stock or from Character Animator and I currently have no license to distribute them. You can purchase the Stock data here:

  • Man: we used all images
  • Dog: we used all images
  • Ostrich: we used the first nine images
  • Cow: we used the first eight images
  • Ape: we used the images from the second, fourth, and the first three images of the fifth row

There are also several websites where you can download Sprite sheets for free.

Train a Model

To train a model with the default parameters from our paper run:

python train.py --gpu_ids 0 --num_keypoints 14 --dataroot datasets/Watercolor-Man --fp16 --name Watercolor-Man

Training one model should take about 60 (FP16) to 90 (FP32) minutes on an NVIDIA GeForce GTX 2080Ti. You can usually use fewer iterations for training and still achieve good results (see next section).

Training Parameters

You can adjust several parameters at train time to possibly improve your results.

  • --name to change the name of the folder in which the results are stored (default is CharacterGAN-Timestamp)
  • --niter 4000 and --niter_decay 4000 to adjust the number of training steps (niter_decayis the number of training steps during which we reduce the learning rate linearly; default is 8000 for both, but you can get good results with fewer iterations)
  • --mask True --output_nc 4 to train with a mask
  • --skeleton False to train without skeleton information
  • --bkg_color 0 to set the background color of the training images to black (default is white, only important if you train with a mask)
  • --batch_size 10 to train with a different batch size (default is 5)

The file options/keypoints.py lets you modify/add/remove keypoints for your characters.

Results

The output is saved to checkpoints/ and we log the training process with Tensorboard. To monitor the progress go to the respective folder and run

 tensorboard --logdir .

Testing

At test time you can either use the model to animate the character or use our interactive GUI to change the position of individual keypoints.

Pretrained Models

Download pretrained models for "Dino", "Fox", "Watercolor-Man", and "Watercolor-Lady" here. Put the model files into the folders in ./checkpoints.

Animate Character

To animate a character (or create interpolations between two images):

python animate_example.py --gpu_ids 0 --model_path checkpoints/Watercolor-Man/ --img_animation_list datasets/Watercolor-Man/animation_list.txt --dataroot datasets/Watercolor-Man

--img_animation_list points to a file that lists the images that should be used for animation. The file should contain one file name per line pointing to an image in dataroot. The model then generates an animation by interpolating between the images in the given order. See datasets/Watercolor-Man/animation_list.txt for an example.

You can add --draw_kps to visualize the keypoints in the animation. You can specifiy the gif parameters by setting --num_interpolations 10 and --fps 5. num_interpolations specifies how many images are generated between two real images (from img_animation_list), fps determines the frames per second of the generated gif.

Use --gpu_ids -1 to run the model on CPU.

Modify Individual Keypoints

To run the interactive GUI:

python visualizer.py --gpu_ids 0 --model_path checkpoints/Watercolor-Man/

Set --gpu_ids -1 to run the model on a CPU. You can also scale the images during visualization, e.g. use --scale 2.

Patch-based Refinement

We use this implementation to run the patch-based refinement step on our generated images. The easiest way to do this is to merge all your training images into a single large image file and use this image file as the style and source image.

Acknowledgements

Our implementation uses code from Pix2PixHD, the TPS augmentation from DeepSIM, and the patch-based refinement code from https://ebsynth.com/ (GitHub).

We would also like to thank Zuzana Studená who produced some of the artwork used in this work. The Dino images were taken from https://www.gameart2d.com.

The data for the human reposing example was obtained from https://endlessreference.com/.

Citation

If you found this code useful please consider citing:

@inproceedings{hinz2022charactergan,
  title={CharacterGAN: Few-Shot Keypoint Character Animation and Reposing},
  author={Hinz, Tobias and Fisher, Matthew and Wang, Oliver and Shechtman, Eli and Wermter, Stefan},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1988--1997},
  year={2022}
}

About

CharacterGAN: Few-Shot Keypoint Character Animation and Reposing (Best Paper WACV 2022)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages