This is a state-of-the-art project for building extraction in high resolution remote sensing image using dataset Massachusetts road & building dataset . And, our approach was published in ACCV 2016, clik here to download our paper

Requirements

• caffe-fcn-master
• OpenCV 2.4.13
• CUDA V8.0
• CUDNN V5.0
• Boost 1.59.0
• Boost.NumPy
• ssai-lib

Boost 1.59.0

tar zxvf boost_1_59_0.tar.gz 
cd boost_1_59_0 
./bootstrap.sh --with-libraries=all --with-toolset=gcc 
./b2 toolset=gcc 
sudo ./b2 install --prefix=/usr 
sudo ldconfig

Boost.NumPy

git clone https://github.com/ndarray/Boost.NumPy.git 
cd Boost.Numpy  
mkdir build 
cd build 
cmake ..   
vim ../CMakeLists.txt   

add some codes before find_package(Boost COMPONENTS Python REQUIRED)  
set(BOOST_ROOT “/usr/include/boost”) 
set(Boost_LIBRARIES “/usr/include/boost/lib”)   
set(Boost_INCLUDE_DIRS “/usr/include/boost/include”) 
set(BOOST_LIBRARYDIR “/usr/include/boost/lib”) 

sudo make 
sudo make install 

ssai-lib

cd ssai-lib/
mkdir build
cd build
cmake ..
make 

Create Dataset

sh shells/download_minh_dataset.sh  
python scripts/create_dataset_256.py  
python scripts/verify_dataset.py -d /data/mass_building/lmdb/train_sat_256 

Start Training

cd models/HF-FCN_Models/BasicNet/  
nohup python solve.py& 

Prediction

cd results/  
python ../scripts/run_prediction.py 
				 --model ../models/HF-FCN_Models/BasicNet/predict.prototxt  
				 --weight ../weights/HF-FCN_iter_12000.caffemodel  
				 --img_dir /data/mass_buildings/source/test/sat  

Evaluation

cd results/prediction_12000   
python ../../scripts/run_evaluation.py   
			--map_dir /data/mass_buildings/source/test/map   
			--result_dir HF-FCN_whole_image_prediction_12000

Results Comparision

whole image comparision

cd shells/
sh run_pr_curve_comparision.sh

	Recall ( \rho = 3 )	Recall ( \rho = 0)	Time (s)
Mnih-CNN \cite{Mnih2013Machine}	0.9271	0.7661	8.70
Mnih-CNN+CRF \cite{Mnih2013Machine}	0.9282	0.7638	26.60
Saito-multi-MA \cite{Saito2016Multiple}	0.9503	0.7873	67.72
Saito-multi-MA&CIS \cite{Saito2016Multiple}	0.9509	0.7872	67.84
Ours (HF-FCN)	0.9643	0.8424	1.07

challenge patches comparision

cd shells/
sh run_recall_comp_for_challenge_patches.sh

Method\Image ID	01	02	03	04	05	06	07	mean
Mnih-CNN+CRF\cite{Mnih2013Machine}	0.784	0.869	0.769	0.653	0.893	0.764	0.800	0.784
Saito-multi-MA&CIS\cite{Saito2016Multiple}	0.773	0.915	0.857	0.789	0.945	0.773	0.830	0.851
Ours (HF-FCN)	0.874	0.964	0.899	0.901	0.986	0.840	0.851	0.911

Pre-trained models

HF-FCN16-iter-12000.caffemodel
Minh13-Machine.caffemodel
Saito16-Multiple-caffemodels

Predicted results

HF-FCN16-results
Mnih13-Machine-results
Saito16-Multiple-results

Reference

If you use this code for your project, please cite this conference paper:
Tongchun Zuo, Juntao Feng, Xuejin Chen. "HF-FCN: Hierarchically Fused Fully Convolutional Network for Robust Building Extraction". Asian Conference of Computer Vision. 2016.