This repository contains the code for the paper OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework [ECCV'2024].

[2024/07] X-Prompt: Our new work X-Prompt: Multimodal Visual Prompt for Video Object Segmentation has been accepted by [ACMMM'2024]. This work proposes a novel multimodal VOS approach, leveraging OneVOS as the RGB-VOS foundation model and incorporating Multi-modal Adaptation Experts to integrate additional modality-specific knowledge. The proposed method achieves SOTA performance across 4 benchmarks.

Our trained models, benckmark scores, and pre-computed results reproduced by this project can be found in MODEL_ZOO.md.

conda create -n vos python=3.9 -y
conda activate vos
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

pip install -r requirements.txt

git clone https://github.com/ClementPinard/Pytorch-Correlation-extension.git
cd Pytorch-Correlation-extension
python setup.py install
cd -

We follow the same data preparation steps used in AOT, including Static datasets, DAVIS and Youtube-VOS. Besides, we also use the more comlicated dataset MOSE and long-term vos dataset LVOS V1 for training or test.

├── OneVOS
├── datasets
│   ├── Static
│   │    └── ...
│   ├── DAVIS
│   │    └── ...
│   ├── YOUTUBE-VOS
│   │    └── ...
│   ├── MOSE
│   │    └── ...
│   ├── LVOS
│   │    └── ...
│   ├── ...

We initialize OneVOS using the weights of ConvMAE-Base as the backbone. You can download the pretrained weights directly from pretrain_backbone and place them in the OneVOS/pretrain_weights/.

Stages:

• PRE: the pre-training stage with static images.
• PRE_YTB_DAV: the main-training stage with YouTube-VOS and DAVIS.
• PRE_YTB_DAV_MOSE: the main-training stage with YouTube-VOS and DAVIS and MOSE.

The training script can be referenced train_examples_pre_ytb_dav.sh, and train_examples_pre_ytb_dav_mose.sh

The inference script can be referenced eval_examples_pre_ytb_dav.sh, and eval_examples_pre_ytb_dav_mose.sh

If you find this repository useful, please consider giving a star and citation:

@inproceedings{li2025onevos,
  title={Onevos: unifying video object segmentation with all-in-one transformer framework},
  author={Li, Wanyun and Guo, Pinxue and Zhou, Xinyu and Hong, Lingyi and He, Yangji and Zheng, Xiangyu and Zhang, Wei and Zhang, Wenqiang},
  booktitle={European Conference on Computer Vision},
  pages={20--40},
  year={2025},
  organization={Springer}
}