Nested UNet (UNet++) for Building Detection

Demo

Input Image	Ground Truth	UNet++ Output	Post-prosessed by Fully Connected CRF	Post-prosessed by Fully Connected CRF and Denoising

Usage

1. Install

   $ git clone https://github.com/hashi0203/Nested-UNet-Building-Detection.git
   $ cd Nested-UNet-Building-Detection
   $ pip install -r requirements.txt

2. Prepare data

    data
    ├── test_img
    │   ├── xxx.png
    │    ︙
    │   └── xxx.png
    ├── train_img
    │   ├── xxx.png
    │    ︙
    │   └── xxx.png
    ├── train_label
    │   ├── xxx.png
    │    ︙
    │   └── xxx.png
    ├── val_img
    │   ├── xxx.png
    │    ︙
    │   └── xxx.png
    └── val_label
        ├── xxx.png
         ︙
        └── xxx.png
    

3. Edit config.py

4. Train

   $ mkdir graph
   $ python train.py

5. Edit config.py

   Set the model to use for evaluation

6. Evaluate

   Warning: Denoising may take a long time

   $ mkdir result
   $ python evaluate.py    # without post processing
   $ python evaluate.py -c # with post processing (Fully Connected CRF)
   $ python evaluate.py -d # with post processing (Fully Connected CRF and denoising)