/motion-cosegmentation

Reference code for "Motion-supervised Co-Part Segmentation" paper

Primary LanguageJupyter Notebook

Motion Supervised co-part Segmentation

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).

Example segmentations

Unsupervised segmentations obtained with our method on VoxCeleb:

and TaiChi dataset:

Example part-swaps

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

Installation

We support python3. To install the dependencies run:

pip install -r requirements.txt

YAML configs

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.

Pre-trained checkpoints

Checkpoints can be found under following links: yandex-disk and google-drive.

Part-swap demo

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 run git 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.

Colab Demo

We prepare a special demo for the google-colab, see: part_swap.ipynb.

Training

Note: It is important to use pytorch==1.0.0 for training. Higher versions of pytorch have strange bilinear warping behavior, because of it model diverge.

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.

Evaluation

We use two metrics to evaluate our model: 1) landmark regression MAE; and 2) Foreground segmentation IoU.

  1. 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
  1. Coming soon...

Datasets

  1. Taichi. Please follow the instruction from https://github.com/AliaksandrSiarohin/video-preprocessing.

  2. VoxCeleb. Please follow the instruction from https://github.com/AliaksandrSiarohin/video-preprocessing.

Training on your own dataset

  1. Follow instructions from First Order Motion Model for preparing your dataset and train First Order Motion Model on your dataset.

  2. This repository use the same dataset format as First Order Motion Model so you can use the same data as in 1).

Additional notes

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}
}