🏆VOTS 2023 Winner: DMAOT（Decoupled Memory AOT）

Intro

DMAOT ranked 1st in the VOTS 2023 challenge (leaderboard). As a plug-and-play method, DMAOT enhances the segmentation ability of AOT series algorithms in long-time videos without requiring any training process.

Instance-wise long-term memories

We decouple the frame-wise long-term memory used in the AOT series frameworks and transform it into instance-wise long-term memory. This enhancement provides more precise control over the long-term memory bank of each individual, facilitating fine-grained memory management.

Dropout frame strategy based on cosine similarity.

We also utilize the dropout frame strategy based on cosine similarity when the maximum number of frames in the instance-wise long-term memory bank is reached. This strategy ensure each long-term memory bank have higher quality of memories.

Prerequisites

Install python packages

• Create a new conda environment

  conda create -n dmaot python=3.8
  conda activate dmaot

• Then run,

  bash install.sh

  Download pretrained model

• Pretrained models of AOT and DeAOT can be downloaded from here.
• We also utilize the SwinB-DeAOT model trained on a larger dataset, and the pretrained weights can be downloaded from here.
• Put the pretrained weight in ./pretrained_models.

  Prepare data

• To initialize the workspace using VOT-Toolkit

  vot initialize vots2023 --workspace <workspace-path>

Run tracker

Edit configuration files

• Edit the paths and env_PATH in trackers.ini.
• Edit the workspace-path in evaluate.sh.
• The detailed documentation on how to use VOT-Toolkit can be found on the VOT Official website.

Get results

bash evaluate.sh

Evaluate

• To zip file using VOT-Toolkit

  vot pack --workspace <workspace-path> <tracker-name>

• Upload to evaluation server. More detail can find in here.

Thanks

DMAOT are based on the AOT-Benchmark, which supports both AOT and DeAOT now. Thanks for such an excellent implementation.

Citations

Please consider citing the related paper(s) in your publications if it helps your research.

@inproceedings{yang2022deaot,
  title={Decoupling Features in Hierarchical Propagation for Video Object Segmentation},
  author={Yang, Zongxin and Yang, Yi},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2022}
}
@article{yang2021aost,
  title={Scalable Video Object Segmentation with Identification Mechanism},
  author={Yang, Zongxin and Wang, Xiaohan and Miao, Jiaxu and Wei, Yunchao and Wang, Wenguan and Yang, Yi},
  journal={arXiv preprint arXiv:2203.11442},
  year={2023}
}
@inproceedings{yang2021aot,
  title={Associating Objects with Transformers for Video Object Segmentation},
  author={Yang, Zongxin and Wei, Yunchao and Yang, Yi},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2021}
}
@inproceedings{kristan2023first,
  title={The first visual object tracking segmentation vots2023 challenge results},
  author={Kristan, Matej and Matas, Ji{\v{r}}{\'\i} and Danelljan, Martin and Felsberg, Michael and Chang, Hyung Jin and Zajc, Luka {\v{C}}ehovin and Luke{\v{z}}i{\v{c}}, Alan and Drbohlav, Ondrej and Zhang, Zhongqun and Tran, Khanh-Tung and others},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={1796--1818},
  year={2023}
}
@article{cheng2023segment,
  title={Segment and Track Anything},
  author={Cheng, Yangming and Li, Liulei and Xu, Yuanyou and Li, Xiaodi and Yang, Zongxin and Wang, Wenguan and Yang, Yi},
  journal={arXiv preprint arXiv:2305.06558},
  year={2023}
}

License

This project is released under the BSD-3-Clause license. See LICENSE for additional details.