/Semantic-Segmentation-with-Sparse-Labels

codes and data for learning from sparse annotations

Primary LanguagePythonMIT LicenseMIT

Semantic-Segmentation-with-Sparse-Labels

The labels and codes for Semantic Segmentation of Remote Sensing Images with Sparse Annotations.

Data

We provid three types of sparse annotations: polygon, scribble, and point. example

Usage

  1. install dependencies in requirements.txt
  2. download and unzip data in the folder data. The directory structure should be as follows:
  path/to/data/
    City/  # Vaihingen or Zurich      
      img/        # images
      line/       # line/scribble-level sparse annotations
      point/      # point-level sparse annotations
      polygon/    # polygon-level sparse annotations
      gt/         # dense gt
      eroded_gt/  # dense gt without boundaries
  1. download and unzip weights in the folder weights.
  2. run python train.py and python test.py for testing and training

Citation

If you find they are useful, please kindly cite the following:

@article{hua2021sparse,
  title={Semantic Segmentation of Remote Sensing Images with Sparse Annotations},
  author={Hua, Yuansheng and Marcos, Diego and Mou, Lichao and Zhu, Xiao Xiang and Tuia, Devis},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={in press}
}