junjue-wang / lovenas

[ISPRS 2024] LoveNAS: Towards Multi-Scene Land-Cover Mapping via Hierarchical Searching Adaptive Network

LoveNAS: Towards Multi-Scene Land-Cover Mapping via Hierarchical Searching Adaptive Network

by Junjue Wang, Yanfei Zhong, Zhuo Zheng, Yuting Wan, Ailong Ma and Liangpei Zhang

[Paper], [Dataset], [BibTeX]

This is an official implementation of LoveNAS.

Environments:

• pytorch >= 1.11.0
• python >=3.6

pip install --upgrade git+https://gitee.com/zhuozheng/[email protected]
pip install git+https://github.com/qubvel/segmentation_models.pytorch
pip install mmcv-full==1.4.7 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html

The Swin-Transformer pretrained weights can be prepared following MMSegmentation.

Search architecture

bash ./scripts/nas_loveda.sh

Train model

The searched architectures and transferred encoder weights should be downloaded.

bash ./scripts/train_loveda.sh

Predict test results

The searched architectures and LoveNAS model weights should be downloaded. Submit the test results to LoveDA Semantic Segmentation Challenge to get scores.

bash ./scripts/submit_loveda.sh

LoveDA

The LoveDA dataset can be downloaded here.

Submit the test results to LoveDA Semantic Segmentation Challenge to get scores.

Search-Config	Backbone	Train-Config	Params (M)	mIoU(%)	Download
config	MobileNetV2	config	3.837	50.60	log&ckpt (pwd:2333)
config	ResNet-50	config	30.491	52.34	log&ckpt (pwd:2333)
config	EfficientNet-B3	config	14.190	52.05	log&ckpt (pwd:2333)
config	Swin-Base	config	92.435	53.76	log&ckpt (pwd:2333)

FloodNet

The FloodNet dataset can be downloaded here.

The train data should be prepared using prepare_floodnet.py.

Search-Config	Backbone	Train-Config	Params (M)	mIoU(%)	Download
config	MobileNetV2	config	12.072	70.73	log&ckpt (pwd:2333)
config	ResNet-50	config	38.457	72.54	log&ckpt (pwd:2333)
config	EfficientNet-B3	config	18.851	72.69	log&ckpt (pwd:2333)
config	Swin-Base	config	97.701	73.79	log&ckpt (pwd:2333)

Citation

If you use LoveNAS in your research, please cite the following papers.

    @article{wang2024lovenas,
      title={LoveNAS: Towards multi-scene land-cover mapping via hierarchical searching adaptive network},
      author={Wang, Junjue and Zhong, Yanfei and Ma, Ailong and Zheng, Zhuo and Wan, Yuting and Zhang, Liangpei},
      journal={ISPRS Journal of Photogrammetry and Remote Sensing},
      volume={209},
      pages={265--278},
      year={2024},
      publisher={Elsevier}
    }
    @inproceedings{wang2021loveda,
        title={Love{DA}: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation},
        author={Junjue Wang and Zhuo Zheng and Ailong Ma and Xiaoyan Lu and Yanfei Zhong},
        booktitle={Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},
        editor = {J. Vanschoren and S. Yeung},
        year={2021},
        volume = {1},
        pages = {},
        url={https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/4e732ced3463d06de0ca9a15b6153677-Paper-round2.pdf}
    }
    @article{wang2020rsnet,
      title={RSNet: The search for remote sensing deep neural networks in recognition tasks},
      author={Wang, Junjue and Zhong, Yanfei and Zheng, Zhuo and Ma, Ailong and Zhang, Liangpei},
      journal={IEEE Transactions on Geoscience and Remote Sensing},
      volume={59},
      number={3},
      pages={2520--2534},
      year={2020},
      publisher={IEEE}
    }

lovenas's Issues

cannot import name 'viz' from 'ever.api'

想请问下代码文件中train_lovada.py中的from ever.api import viz，报错提示无法导入viz，这个要怎样解决呢？
我尝试了ever-bata-0.2.3，0.4.1。还是需要哪个特定的版本呢？
我同样尝试了pip install --upgrade git+https://gitee.com/zhuozheng/[email protected]，但是这个没有起作用，我直接打开链接https://gitee.com/zhuozheng/ever_lite显示我的访问受限。
请问该怎样解决呢。

权重

你好，需要的权重在哪下（例如resnet50_loveda_30k.pth）

Holy god

🐢🐢, this work will undoubtedly be one of my favourites this year!