This repository contains the source code for the CVPR oral paper Animating Arbitrary Objects via Deep Motion Transfer by Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci and Nicu Sebe. We call the proposed deep framework Monkey-Net, as it enables motion transfer by considering MOviNg KEYpoints. Check also the project website.
The videos on the left show the driving videos. The first row on the right for each dataset shows the source images. The bottom row contains the animated sequences with motion transferred from the driving video and object taken from the source image. We trained a separate network for each task. Note that for each task the background, the object appearance are consistent in each generated video.
Our framework can be used in several modes. In the motion transfer mode, a static image will be animated using a driving video. In the image-to-video translation mode, given a static image, the framework will predict future frames.
We support python3. To install the dependencies run:
pip install -r requirements.txt
There are several configuration (config/dataset_name.yaml) files one for each dataset. See config/actions.yaml to get description of each parameter.
To run a demo, download a checkpoint and run the following command:
python demo.py --config config/moving-gif.yaml --driving_video sup-mat/driving.png --source_image sup-mat/source.png --checkpoint path/to/checkpoint
The result will be stored in demo.gif.
To train a model on specific dataset run:
CUDA_VISIBLE_DEVICES=0 python run.py --config config/dataset_name.yaml
The code will create a folder in the log directory (each run will create a time-stamped new directory).
Checkpoints will be saved to this folder.
To check the loss values during training in see log.txt.
You can also check training data reconstructions in the train-vis subfolder.
To evaluate the reconstruction performance run:
CUDA_VISIBLE_DEVICES=0 python run.py --config config/dataset_name.yaml --mode reconstruction --checkpoint path/to/checkpoint
You will need to specify the path to the checkpoint,
the reconstruction subfolder will be created in the checkpoint folder.
The generated video will be stored to this folderenerated video there and in png subfolder loss-less verstion in '.png' format.
In order to perform motion transfer run:
CUDA_VISIBLE_DEVICES=0 python run.py --config config/dataset_name.yaml --mode transfer --checkpoint path/to/checkpoint
You will need to specify the path to the checkpoint,
the transfer subfolder will be created in the same folder as the checkpoint.
You can find the generated video there and its loss-less version in the png subfolder.
There are 2 different ways of performing transfer: by using absolute keypoint locations or by using relative keypoint locations.
-
Absolute Transfer: the transfer is performed using the absolute postions of the driving video and appearance of the source image. In this way there are no specific requirements for the driving video and source appearance that is used. However this usually leads to poor performance since unrelevant details such as shape is transfered. Check transfer parameters in
shapes.yamlto enable this mode. -
Realtive Transfer: from the driving video we first estimate the relative movement of each keypoint, then we add this movement to the absolute position of keypoints in the source image. This keypoint along with source image is used for transfer. This usually leads to better performance, however this requires that the object in the first frame of the video and in the source image have the same pose.
The approximately aligned pairs of videos are given in the data folder. (e.g data/taichi.csv).
In order to perform image-to-video translation run:
CUDA_VISIBLE_DEVICES=0 python run.py --config config/dataset_name.yaml --mode prediction --checkpoint path/to/checkpoint
The following steps will be performed:
- Estimate the keypoints from the training set
- Train rnn to predict the keypoints
- Run the predictor for each video in the dataset, starting from the first frame.
Again the
predictionsubfolder will be created in the same folder as the checkpoint. You can find the generated video there and inpngsubfolder.
-
Shapes. This dataset is saved along with repository. Training takes about 1 hour.
-
Actions. This dataset is also saved along with repository. And training takes about 4 hours.
-
Nemo. The preprocessed version of this dataset can be downloaded. Training takes about 6 hours.
-
Taichi. We used the same data as MoCoGAN. Training takes about 15 hours.
-
Bair. The preprocessed version of this dataset can be downloaded. Training takes about 4 hours.
-
MGif. The preprocessed version of this dataset can be downloaded. Check for details on this dataset. Training takes about 8 hours, on 2 gpu.
-
Vox. The dataset can be downloaded and preprocessed using a script:
cd data; ./get_vox.sh.
-
Resize all the videos to the same size e.g 128x128, the videos can be in '.gif' or '.mp4' format. But we recommend to make them stacked '.png' (see data/shapes), because this format is lossless.
-
Create a folder
data/dataset_namewith 2 subfolderstrainandtest, put training videos in thetrainand testing in thetest. -
Create a config
config/dataset_name.yaml(it is better to start from one of the existing configs, for 64x64 videosconfig/nemo.yaml, for 128x128config\moving-gif.yaml, for 256x256config\vox.yaml), in dataset_params specify the root dir theroot_dir: data/dataset_name. Also adjust the number of epoch in train_params.
Citation:
@InProceedings{Siarohin_2019_CVPR,
author={Siarohin, Aliaksandr and Lathuilière, Stéphane and Tulyakov, Sergey and Ricci, Elisa and Sebe, Nicu},
title={Animating Arbitrary Objects via Deep Motion Transfer},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}