Skip to content

Latest commit

 

History

History
55 lines (44 loc) · 2.28 KB

README.md

File metadata and controls

55 lines (44 loc) · 2.28 KB

Requirements

  • Python3 (tested with 3.5.4)
  • GDAL (tested with 2.2.2)
  • Numpy (tested with 1.14.2)
  • Pillow (tested with 5.0.0)
  • OpenCV (tested with 3.3.1)
  • MulticoreTSNE (tested with 0.0.1.1)
  • colorcorrect (tested with 0.7)
  • tqdm (tested with 4.15.0)
  • PyTorch (tested with 0.4.1)
  • TorchVision (tested with 0.2.1)

Usage

Convert RGB-NIR GeoTIFF data into cropped RGB and NIR images

crop_rgb-nir.py separates RGB-NIR GeoTIFF data into RGB and NIR data and crops the each data except for blackout part.
If filename option is specified, it is processed for the specified file with single core cpu.
If not, it is processed for files in the specified input directory with multicore cpu.
If you specified --colorcorrect (-cc) option, cropped images are colorcorrected using colorcorrect module.

python crop_rgb-nir.py -i <path_to_input_dir> -o <path_to_output_dir> --filename <filename> -s <crop_size> -cc

or

python crop_rgb-nir.py -i <path_to_input_dir> -o <path_to_output_dir> -s <crop_size> -cc

Making a list file of training data

make_training_datalist.py makes a list of training data train_files.pkl from cropped RGB images.
make_training_datalist.py saves feature vectors of fc7 layer of pretrained AlexNet as an intermediate result into filename_feature.pkl.
For details, refer to Sec. 3.2 of our paper.

python make_training_datalist.py -i <path_to_input_dir_or_filename_feature.pkl> -o <path_to_output_dir> -n_d <num_of_training_data> -n_g <square_of_num_of_grids>

Synthesis of cloud images

make_clouds.py makes synthesized cloud images using perlin noise.
The size of the cloud can be adjusted by changing NoiseOffset in the PythonCloud/Config.py.

python make_clouds.py -n <num_of_cloud_images> -o <path_to_output_dir>

Visualizing a feature space (optional)

feature_space_visualizer.py make an image visualized 2-D feature space from filename_feature.pkl.

python feature_space_visualizer.py -i <path_to_filename_feature.pkl> -o <path_to_output_file> -g_n <square_of_num_of_grids>

References

  • PythonCloud/ is refered Python-Cloud repository. We modified the code of this repository for Python3.