Towards Temporally Consistent Referring Video Object Segmentation

Introduction

Referring Video Object Segmentation (R-VOS) methods face challenges in maintaining consistent object segmentation due to temporal context variability and the presence of other visually similar objects. We propose an end-to-end R-VOS paradigm that explicitly models temporal instance consistency alongside the referring segmentation. Furthermore, we propose a new Mask Consistency Score (MCS) metric to evaluate the temporal consistency of video segmentation. Extensive experiments demonstrate that our approach enhances temporal consistency by a significant margin, leading to top-ranked performance on popular R-VOS benchmarks.

https://github.com/bo-miao/HTR/assets/53172019/7b2e7d56-59f8-4ba2-b502-c4e7ed9e0417

Installation and Data Preparation

Please refer to SgMg for installation and data preparation.

Evaluation

The checkpoint for HTR w/ SwinL is available at HTR-SwinL.

If you want to evaluate HTR on Ref-DAVIS/YouTube-VOS, please run the following command in the scripts folder:

sh dist_test_davis_swinl.sh

sh dist_test_ytv_swinl.sh

MCS Metric for Temporal Consistency

The code for MCS evaluation is in get_mcs.py. Please click View scoring output log to download stdout.txt of your submission in Ref-YTVOS eval server.

Then you can run the script to get the MCS score under different thresholds.

Citation

@article{miao2024htr,
  title={Towards Temporally Consistent Referring Video Object Segmentation},
  author={Miao, Bo and Bennamoun, Mohammed and Gao, Yongsheng and Shah, Mubarak and Mian, Ajmal},
  journal={https://arxiv.org/abs/2403.19407},
  year={2024}
}

Acknowledgements

Contact

If you have any questions about this project, please feel free to contact bomiaobbb@gmail.com.

bo-miao / HTR

readme