SI-LACMMT

Level-Aware Consistent Multi-level Map Translation From Satellite Imagery

Ying Fu, Zheng Fang, Linwei Chen, Tao Song, and Defu Lin

1. Image2image translation and map generation methods

Method	FID$\downarrow$	KMMD$\downarrow$	WD$\downarrow$	PSNR$\uparrow$	Model
Pix2pix	342.84	0.53236	16.267	20.726	Baidu Disk
Pix2pixHD	331.10	0.38806	14.358	20.908	Baidu Disk
CycleGAN	312.14	0.45603	14.692	20.725	Baidu Disk
SPADE	459.11	0.79792	20.047	20.468	Baidu Disk
SelectionGAN	337.83	0.58850	16.475	20.617	Baidu Disk
TSIT	284.17	0.43861	13.753	20.540	Baidu Disk
LPTN	351.61	0.4380	17.489	21.327	Baidu Disk
SMAPGAN	336.35	0.54771	16.338	$\bf{22.506}$	Baidu Disk
CreativeGAN	267.37	0.29931	12.453	21.428	Baidu Disk
SI-LACMMT	$\bf{195.64}$	$\bf{0.23759}$	$\bf{11.014}$	21.532	Baidu Disk
LAMG_MLMG	-	-	-	-	Baidu Disk

2. Create Environment:

Python 3 (Recommend to use Anaconda)
NVIDIA GPU + CUDA
Python packages:

pip install -r requirements.txt

3. MLMG Dataset:

The advantages of our dataset over other datasets are as follows：

Our data contains multiple types of images, and has multi-level data in the same area. In addition, we have balanced the data of each level to ensure the balance of training.

The specific content of our data set is as shown above, including data from US and CN countries. The data volume of each level of training set is 2k. The test set is multi-level data in the same region, and the number of high-level data is four times that of low-level data.

following form:

|--MLMG Dataset
    |--US_dataset
    	|-- trainA(SI images)
    		|-- 15
    			|-- 15_26967_12413.jpg
    			|-- 15_26967_12415.jpg
    			：
                |-- 15_27091_12529.jpg
    		|-- 16
    		|-- 17
    		|-- 18
    	|-- trainB(Map images)
    	|-- trainC(2x downsample map images )
    	|-- train_seg(element labels)
    	|-- testA(SI images)
    	|-- testB(Map images)
    |--CN_dataset
    	|-- trainA
    	|-- trainB
    	|-- trainC
    	|-- train_seg
    	|-- testA
    	|-- testB

4. Experiement:

4.1 Training

Download the dataset and move it to the dataset folder.
Download the pretrained model and move it to src/LACMMT/.
Execute the following command to train the corresponding model.

cd "root_path"
bash scripts/train_CreativeGAN.sh
bash scripts/train_LACMMT.sh
bash scripts/train_LAMG_MLMG.sh

4.2 Testing

Download the test dataset and the pretrained model provided above.
Configure the python environment according to the requests.txt.
Download the dataset and move it to the dataset folder.
Execute the following script to generate the multi-level map images.

cd "root_path"
bash scripts/test_LACMMT.sh
bash scripts/test_Pix2pix.sh
:
bash scripts/test_CreativeGAN.sh

5.Citation

@article{fu2022level,
  title={Level-Aware Consistent Multilevel Map Translation From Satellite Imagery},
  author={Fu, Ying and Fang, Zheng and Chen, Linwei and Song, Tao and Lin, Defu},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={61},
  pages={1--14},
  year={2022},
  publisher={IEEE}
}
@article{chen2022consistency,
  title={Consistency-Aware Map Generation at Multiple Zoom Levels Using Aerial Image},
  author={Chen, Linwei and Fang, Zheng and Fu, Ying},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  volume={15},
  pages={5953--5966},
  year={2022},
  publisher={IEEE}
}
@article{FZ_transformer,
  title={Transformer特征引导的双阶段地图智能生成},
  author={方政 and 付莹 and 刘利雄}
}

6.News ✨

2023-1-1: release train code for CreativeGAN, LAMG_MLMG, and LACMMT.
2022-12-18: release test code.
2022-8-30: release dataset.

FZfangzheng/SI-LACMMT