DiffusionInst: Diffusion Model for Instance Segmentation

DiffusionInst is the first work of diffusion model for instance segmentation. We hope our work could serve as a simple yet effective baseline, which could inspire designing more efficient diffusion frameworks for challenging discriminative tasks.

DiffusionInst: Diffusion Model for Instance Segmentation
Zhangxuan Gu, Haoxing Chen, Zhuoer Xu, Jun Lan, Changhua Meng, Weiqiang Wang arXiv 2212.02773

Todo list:

[x] Release source code.
[x] Hyper-paramters tuning.
[x] Add Swin-Large backbone.
[x] Release trained models.
[ ] Adding directly filter denoising.

Getting Started

The installation instruction and usage are in Getting Started with DiffusionInst.

Trained Models

We now provide trained models for ResNet-50 and ResNet-101.

https://pan.baidu.com/s/1KEdjNY3CSXWp0VFwkhRKYg, pwd: jhbv.

Model Performance

Method | Mask AP (1 step) | Mask AP (4 step) --- |:---:|:---: COCO-val-Res50 | 37.3| 37.5 COCO-val-Res101 | 41.0| 41.1 COCO-val-Swin-B| 46.6| 46.8 COCO-val-Swin-L| 47.8| 47.8 LVIS-Res50 | 22.3| - LVIS-Res101| 27.0| - LVIS-Swin-B| 36.0| - COCO-testdev-Res50 | 37.1| - COCO-testdev-Res101 | 41.5| - COCO-testdev-Swin-B| 47.6| - COCO-testdev-Swin-L| 48.3| -

Star History

Citing DiffusionInst

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

@article{DiffusionInst,
      title={DiffusionInst: Diffusion Model for Instance Segmentation},
      author={Gu, Zhangxuan and Chen, Haoxing and Xu, Zhuoer and Lan, Jun and Meng, Changhua and Wang, Weiqiang},
      journal={arXiv preprint arXiv:2212.02773},
      year={2022}
}

Acknowledgement

Many thanks to the nice work of DiffusionDet @ShoufaChen. Our codes and configs follow DiffusionDet.

Contacts

Please feel free to contact us if you have any problems.

Email: haoxingchen@smail.nju.edu.cn or guzhangxuan.gzx@antgroup.com

DiffusionInst

Install / Use

README