VANESSA: VAihiNgEn Semantic SegmentAtion
This was tested in a system with GPU NVIDIA GeForce GTX 1070 and Ubuntu 16.04.7.
- You need to have installed the drivers of your GPU if you are using one.
- You need to create a folder called 'vaihingen/', and create there two folders:
- 'images/': And copy there all the images (top_mosaic_09cm_area1.tif, ..., top_mosaic_09cm_area38.tif)
- 'labels/': And copy there all the labels (top_mosaic_09cm_area1.tif, ..., top_mosaic_09cm_area38.tif)
- Also, you need to copy the included 'sets.csv' file, to the 'vaihingen/' folder.
- Then, you need to edit the file "segmentation/training_pixel_tile_seg_model.py", configuring the variable 'path', in the line 115 to the path of the folder 'vaihingen/', for example: 'path = "/home/sebastian/vaihingen"'.
- The script "install.sh" creates a python virtual environment and installs in it all the dependencies in the requirements.txt file.
- The script "run_tests.sh" activates the virtual environment and runs the tests of the project.
- The script "train_pixel_tile_level.sh" activates the virtual environment
and runs the python script "segmentation/training_pixel_tile_seg_model.py",
which runs two models to compare between them:
- A pixel level segmentation model that uses a U-Net like structure.
- A pixel level segmentation model that uses a U-Net like structure, but with a second output for pretraining with patch level data, to be able to leverage image data with a more coarse labeled information.