PDNet

Towards imbalanced motion: part-decoupling network for video portrait segmentation (SCIS 2024)

[Paper] [arXiv] [中文介绍]

MVPS dataset

To acquire our MVPS dataset, please contact yuts@buaa.edu.cn via an official email. Please explain the purpose of usage in the email. The dataset is available for academic research only.

evaluation code

Image portrait segmentation datasets for training

The datasets are from EG1800, Supervisely Person and MS COCO.

download link

Model weights

ResNet-50 weights

ResNet-101 weights

When using ResNet-101 as backbone, please set the MODEL.BACKBONE parameter in config.yaml to 'resnet101'.

Command for training/test

PyTorch version: 1.11.0

Training: torchrun --nproc_per_node=2 runner.py

Test: python runner.py --test

Citation

@article{PDNet,
  author={Tianshu Yu and Changqun Xia and Jia Li},
  title={Towards imbalanced motion: part-decoupling network for video portrait segmentation},
  journal={SCIENCE CHINA Information Sciences},
  year={2024},
  volume={67},
  number={7},
  pages={172104},
  doi={10.1007/s11432-023-4030-y}
}