This is the official code for the paper:

End-to-End Learning of Polygons for Remote Sensing Image Classification
Nicolas Girard, Yuliya Tarabalka
IGARSS 2018
[Paper] [Slides]

• Tensorflow 1.4

1. Train the polygon Encoder-Decoder network. This is used to pre-train the weights of the Decoder part of PolyCNN. See the corresponding subdirectory.
2. Download and setup the "Distributed Solar Photovoltaic Array Location and Extent Data Set for Remote Sensing Object Identification" dataset, see the corresponding subdirectory.
3. Download the pre-trained InceptionV4 checkpoint, see the corresponding subdirectory.
4. Train PolyCNN and run inference on the test set, see the corresponding subdirectory.
5. Train the U-Net of unet_and_vectorization and run inference on the test set, see the corresponding subdirectory.
6. Compare the two methods, see the corresponding subdirectory.

@inproceedings{girard:hal-01762446,
  TITLE = {{End-to-End Learning of Polygons for Remote Sensing Image Classification}},
  AUTHOR = {Girard, Nicolas and Tarabalka, Yuliya},
  URL = {https://hal.inria.fr/hal-01762446},
  BOOKTITLE = {{IEEE International Geoscience and Remote Sensing Symposium -- IGARSS 2018}},
  ADDRESS = {Valencia, Spain},
  YEAR = {2018},
  MONTH = Jul,
  KEYWORDS = {convolutional neural networks ; Index Terms- High-resolution aerial images ;  polygon ; vectorial ;  regression ;  deep learning},
  PDF = {https://hal.inria.fr/hal-01762446/file/girard.pdf},
  HAL_ID = {hal-01762446},
  HAL_VERSION = {v1},
}