🚀🚀🚀 Official implementation of SeC: Advancing Complex Video Object Segmentation via Progressive Concept Construction
Zhixiong Zhang* ·
Shuangrui Ding* ·
Xiaoyi Dong✉ ·
Songxin He ·
Jianfan Lin ·
Junsong Tang
Yuhang Zang ·
Yuhang Cao ·
Dahua Lin ·
Jiaqi Wang✉
demo.mp4
🚀 [2025/12/22] Training code has been fully open-sourced.
🚀 [2025/10/14] SeC is cited by SAM 3 and used as a baseline!
🏆 [2025/8/13] SeC sets a new state-of-the-art on the latest MOSE v2 leaderboard!
🚀 [2025/7/22] The Paper and Project Page are released!
- 🔥We introduce Segment Concept (SeC), a concept-driven segmentation framework for video object segmentation that integrates Large Vision-Language Models (LVLMs) for robust, object-centric representations.
- 🔥SeC dynamically balances semantic reasoning with feature matching, adaptively adjusting computational efforts based on scene complexity for optimal segmentation performance.
- 🔥We propose the Semantic Complex Scenarios Video Object Segmentation (SeCVOS) benchmark, designed to evaluate segmentation in challenging scenarios.
| Model | SA-V val | SA-V test | LVOS v2 val | MOSE v1 | MOSE v2 | DAVIS 2017 val | YTVOS 2019 val | SeCVOS |
|---|---|---|---|---|---|---|---|---|
| SAM 2.1 | 78.6 | 79.6 | 84.1 | 74.5 | 49.5 | 90.6 | 88.7 | 58.2 |
| SAMURAI | 79.8 | 80.0 | 84.2 | 72.6 | 51.1 | 89.9 | 88.3 | 62.2 |
| SAM2.1Long | 81.1 | 81.2 | 85.9 | 75.2 | 51.5 | 91.4 | 88.7 | 62.3 |
| SeC (Ours) | 82.7 | 81.7 | 86.5 | 75.3 | 53.8 | 91.3 | 88.6 | 70.0 |
- Release SeC training code
- Release SeCVOS benchmark annotations
- Release SeC inference code and checkpoints
Please make sure using the correct versions of transformers and peft.
conda create -n sec python=3.10
conda activate sec
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txtDownload the SeC checkpoint from 🤗HuggingFace and place it in the following directory :
saved_models
├── SeC-4B
│ └── config.json
│ └── generation_config.json
...
If you want to test SeC inference on a single video, please refer to demo.ipynb.
The inference instruction is in INFERENCE.md.
The evaluation instruction can be found in EVALUATE.md. To evaluate performance on seen and unseen categories in the LVOS dataset, refer to the evaluation code available here.
Training instructions can be found in TRAIN.md.
To train SeC on your own dataset, simply organize your data following the format described there.
This repository are licensed under a Apache License 2.0.
This repo benefits from SAM 2, SAM2Long and Sa2VA. Thanks for their wonderful works.
If you find our work helpful for your research, please consider giving a star ⭐ and citation 📝
@article{zhang2025sec,
title = {SeC: Advancing Complex Video Object Segmentation via Progressive Concept Construction},
author = {Zhixiong Zhang and Shuangrui Ding and Xiaoyi Dong and Songxin He and Jianfan Lin and Junsong Tang and Yuhang Zang and Yuhang Cao and Dahua Lin and Jiaqi Wang},
journal = {arXiv preprint arXiv:2507.15852},
year = {2025}
}