This repository contains the source code for the paper Motion Supervised co-part Segmentation by Aliaksandr Siarohin*, Subhankar Roy*, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci and Nicu Sebe.
* - denotes equal contribution
Our method is a self-supervised deep learning method for co-part segmentation. Differently from previous works, our approach develops the idea that motion information inferred from videos can be leveraged to discover meaningful object parts. Our method can also perform video editing (aka part-swaps).
Unsupervised segmentations obtained with our method on VoxCeleb:
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
and TaiChi dataset:
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
Part swaping with our method for VoxCeleb dataset. Each triplet shows source image, target video (with swap mask in the corner) and result:
Hair Swap | Beard Swap | ||||
---|---|---|---|---|---|
![]() |
![]() |
||||
Eyes Swap | Lips Swap | ||||
![]() |
![]() |
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/taichi-sem-256.yaml
to get description of each parameter.
Checkpoints can be found under following links: yandex-disk and google-drive.
To run a demo, download checkpoint and run the following command:
python part_swap.py --config config/dataset_name.yaml --target_video path/to/target --source_image path/to/source --checkpoint path/to/checkpoint --swap_index 0,1
The result will be stored in result.mp4
.
-
For swaping either soft or hard labels can be used (specify
--hard
for hard segmentation). -
For swaping either target or source segmentation mask can be used (specify
--use_source_segmentation
for using source segmentation mask). -
For the reference we also provide fully-supervised segmentation. For fully-supervised add
--supervised
option. And rungit clone https://github.com/AliaksandrSiarohin/face-makeup.PyTorch face_parsing
which is a fork of @zllrunning. -
Also for the reference we provide First Order Motion Model based alignment, use
--first_order_motion_model
and the correspoinding checkpoint. This allignment can only be used along with--supervised
option.
We prepare a special demo for the google-colab, see: part_swap.ipynb
.
Model training consist in finetuning the First Order Model checkpoint (they can be downloaded from google-drive or yandex-disk). Use the following command for training:
CUDA_VISIBLE_DEVICES=0 python train.py --config config/dataset_name.yaml --device_ids 0 --checkpoint dataset-name.cpk.pth.tar
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.
By default the batch size is tunned to run on 1 Tesla-p100 gpu, you can change it in the train_params in the corresponding .yaml
file.
We use two metrics to evaluate our model: 1) landmark regression MAE; and 2) Foreground segmentation IoU.
- For computing the MAE download eval_images.tar.gz from google-drive-eval and use the following command:
CUDA_VISIBLE_DEVICES=0 python evaluate.py --config config/dataset_name.yaml --device_ids 0 --root_dir path-to-root-folder-of-dataset --checkpoint_path dataset-name.cpk.pth.tar
- Coming soon...
-
Taichi. Please follow the instruction from https://github.com/AliaksandrSiarohin/video-preprocessing.
-
VoxCeleb. Please follow the instruction from https://github.com/AliaksandrSiarohin/video-preprocessing.
-
Follow instructions from First Order Motion Model for preparing your dataset and train First Order Motion Model on your dataset.
-
This repository use the same dataset format as First Order Motion Model so you can use the same data as in 1).
Citation:
Motion Supervised co-part Segmentation:
@article{Siarohin_2020_motion,
title={Motion Supervised co-part Segmentation},
author={Siarohin, Aliaksandr and Roy, Subhankar and Lathuilière, Stéphane and Tulyakov, Sergey and Ricci, Elisa and Sebe, Nicu},
journal={arXiv preprint},
year={2020}
}
First Order Motion Model:
@InProceedings{Siarohin_2019_NeurIPS,
author={Siarohin, Aliaksandr and Lathuilière, Stéphane and Tulyakov, Sergey and Ricci, Elisa and Sebe, Nicu},
title={First Order Motion Model for Image Animation},
booktitle = {Conference on Neural Information Processing Systems (NeurIPS)},
month = {December},
year = {2019}
}