thera0810 / boxlevelset

The code for "Box-supervised Instance Segmentation with Level Set Evolution" (ECCV2022).

Box-supervised Instance Segmentation with Level Set Evolution

Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, Xiansheng Hua, Lei Zhang

Paper (Arxiv)

Installation

This implementation is based on MMdetection. Please refer to install.md for detailed installation.

Getting Started

Please see getting_started.md for models training and inference.

Performance

Models

• The following models are trained with Telsa V100 GPU.
• The pretrained models are in GoogleDriver.

Mask AP Results on Pascal VOC val

Backbone	schd	Models	GPUs	AP	AP_25	AP_50	AP_70	AP_75
ResNet-50	3x	model	4	36.5	76.8	64.2	44.8	36.4
ResNet-101	3x	model	4	38.3	77.9	66.3	46.4	38.7

Mask AP Results on COCO 2017

Backbone	schd	Models	GPUs	AP(val)	AP(test-dev)
ResNet-50	3x	model	8	31.4	31.7
ResNet-101	3x	model	8	33.0	33.4
ResNet-101-DCN	3x	model	8	35.0	35.4

Note:

• Following BBTP and DiscoBox, the Pascal VOC is aumented Pascal VOC(data link) with SBD. We recomment the users to train the Pascal VOC first to validate the performance with ~14 hours training time.
• Training COCO with 3x needs about 4 days.

Mask AP Results on iSAID val

Backbone	schd	Models	GPUs	input	AP	AP_50	AP_75
ResNet-50	1x	model	4	800*800	24.3	48.1	20.7

Note:

• The high-resolution images of isaid are splitted into 800*800 patches.
• The input size of network is also set to 800*800.
• The train subset is used for model training, the val subset is for performance evaluation.

Visual Results on General Scene

Visual Results on iSAID (remote sensing)

• The bounding boxes are generated by the mask predictions.

Please run the following script to get more visual results.

   python tools/test.py configs/boxlevelset/config-xxx.py work_dirs/xxx.pth  --show-dir show_dirs/

Citation

@article{li2022boxlevelset,
  title={Box-supervised Instance Segmentation with Level Set Evolution},
  author={Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, Xiansheng Hua, Lei Zhang},
  journal={ECCV2022},
  year={2022}
}

Acknowledgements

SOLO

AdelaiDet

License

For academic use, this project is licensed under the Apache License 2.0. For commercial use, please contact the authors.