Building Detection from High Resolution Satellite Images

Implementation of Fully Convolutional Network, U-Net, Deep Residual U-Net, Pyramid Scene Parsing Network and Deep Structured Active Contour.

Papers

Fully Convolutional Network Paper: Link
SegNet Paper: Link
U-Net Paper: Link
Deep UNet or Residual UNet Paper: Link
PSPNet Paper: Link

Architectures

Dataset

The datasets used in this project can be downloaded from the following links

Inria Building Dataset:

ISPRS Dataset:

SpaceNet Challenge Dataset:

CrowdAI Mapping Challenge Dataset:

WHU Building Dataset:

WHU East Asia Dataset:

WHU New Zealand:

WHU Global Cities:

Requirements

• 'OpenCV-Python'
• Matplotlib
• Numpy
• Scikit Images
• Keras
• Tensorflow
• GDAL
• PyCOCOTools

Usage

Help

python3 main.py --help  

Split Dataset

python3 main.py -sd True -imp ../../Dataset/Inria_Dataset/train/ -od ../../Dataset/Inria_Patches/

Train a Model

python3 main.py -pp ../../Dataset/Inria_Patches/inria_dataset_256/train/ -t True -m unet

Training Plots

Results

Without Augmentations

Places	FCN	SegNet	U-Net	Deep U-Net	PSPNet
Austin	0.43	0.52	0.55	0.52	0.53
Chicago	0.59	0.60	0.73	0.62	0.65
Kitsap	0.23	0.29	0.37	0.41	0.44
Tyrol	0.19	0.26	0.31	0.29	0.43
Vienna	0.51	0.48	0.77	0.73	0.63

After Augmentations

Places	FCN	SegNet	U-Net	Deep U-Net	PSPNet
Austin	0.54	0.57	0.69	0.63	0.60
Chicago	0.63	0.63	0.79	0.67	0.67
Kitsap	0.44	0.51	0.59	0.58	0.53
Tyrol	0.54	0.56	0.70	0.69	0.60
Vienna	0.68	0.72	0.80	0.77	0.68