BoxVIS: Video Instance Segmentation with Box Annotation

Minghan LI and Lei ZHANG

[arXiv]

Updates

• July 13, 2023: Paper has been updated.
• June 30, 2023: Code and trained models are available now.
• March 28, 2023: Paper is available now.

Installation

See installation instructions.

Datasets

See Datasets preparation.

Getting Started

We provide a script train_net_boxvis.py, that is made to train all the configs provided in BoxVIS.

Training: download pretrained weights of Mask2Former and save it into the path 'pretrained/*.pth', then run:

sh run.sh

Inference: download trained weights, and save it into the path 'pretrained/*.pth', then run:

sh test.sh

Quantitative performance comparison

Citing BoxVIS

If you use BoxVIS in your research or wish to refer to the baseline results published in the Model Zoo, please use the following BibTeX entry.

@misc{li2023boxvis,
      title={BoxVIS: Video Instance Segmentation with Box Annotations}, 
      author={Minghan Li and Lei Zhang},
      year={2023},
      eprint={2303.14618},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

Our code is largely based on Detectron2, Mask2Former, MinVIS, and VITA. We are truly grateful for their excellent work.