HS2S

Pytorch code for Hybrid-S2S: Video Object Segmentation with Recurrent Networks and Correspondence Matching
For more information, please visit our project page
Dependencie:

numpy==1.18.1,
sacred==0.8.1,
torch==1.5,
torchvision==0.6,
tqdm

For training on Youtube-VOS, download the dataset from here and modify the data paths in train.py. Finally, run the following command.

python train.py

For inference with the pretrained model, download the weights from here and put them under the Model directory.

Modify the configurations and data paths in submit_ytvos.py and run the following command.

python submit_ytvos.py with model_name='weights_HS2S.pth'

In case of questions, please contact fatemeh.azimi@dfki.de and if this repository is useful for you work, please consider citing our paper:

@article{azimi2020hybrid,
  title={Hybrid Sequence to Sequence Model for Video Object Segmentation},
  author={Azimi, Fatemeh and Frolov, Stanislav and Raue, Federico and Hees, Joern and Dengel, Andreas},
  journal={arXiv preprint arXiv:2010.05069},
  year={2020}
}