Awesome

This is an implementation of Volodymyr Mnih's dissertation methods on his Massachusetts road & building dataset and my original methods that are published in this paper.

Requirements

• Python 3.5 (anaconda with python 3.5.1 is recommended)
  • Chainer 1.5.0.2
  • Cython 0.23.4
  • NumPy 1.10.1
  • tqdm
• OpenCV 3.0.0
• lmdb 0.87
• Boost 1.59.0
• Boost.NumPy (26aaa5b)

Build Libraries

OpenCV 3.0.0

$ wget https://github.com/Itseez/opencv/archive/3.0.0.zip
$ unzip 3.0.0.zip && rm -rf 3.0.0.zip
$ cd opencv-3.0.0 && mkdir build && cd build
$ bash $SSAI_HOME/shells/build_opencv.sh
$ make -j32 install

If some libraries are missing, do below before compiling 3.0.0.

$ sudo apt-get install -y libopencv-dev libtbb-dev

Boost 1.59. 0

$ wget http://downloads.sourceforge.net/project/boost/boost/1.59.0/boost_1_59_0.tar.bz2
$ tar xvf boost_1_59_0.tar.bz2 && rm -rf boost_1_59_0.tar.bz2
$ cd boost_1_59_0
$ ./bootstrap.sh
$ ./b2 -j32 install cxxflags="-I/home/ubuntu/anaconda3/include/python3.5m"

Boost.NumPy

$ git clone https://github.com/ndarray/Boost.NumPy.git
$ cd Boost.NumPy && mkdir build && cd build
$ cmake -DPYTHON_LIBRARY=$HOME/anaconda3/lib/libpython3.5m.so ../
$ make install

Build utils

$ cd $SSAI_HOME/scripts/utils
$ bash build.sh

Create Dataset

$ bash shells/download.sh
$ bash shells/create_dataset.sh

Start Training

$ CHAINER_TYPE_CHECK=0 CHAINER_SEED=$1 \
nohup python scripts/train.py \
--seed 0 \
--gpu 0 \
--model models/MnihCNN_multi.py \
--train_ortho_db data/mass_merged/lmdb/train_sat \
--train_label_db data/mass_merged/lmdb/train_map \
--valid_ortho_db data/mass_merged/lmdb/valid_sat \
--valid_label_db data/mass_merged/lmdb/valid_map \
--dataset_size 1.0 \
> mnih_multi.log 2>&1 < /dev/null &

Prediction

python scripts/predict.py \
--model results/MnihCNN_multi_2016-02-03_03-34-58/MnihCNN_multi.py \
--param results/MnihCNN_multi_2016-02-03_03-34-58/epoch-400.model \
--test_sat_dir data/mass_merged/test/sat \
--channels 3 \
--offset 8 \
--gpu 0 &

Evaluation

$ PYTHONPATH=".":$PYTHONPATH python scripts/evaluate.py \
--map_dir data/mass_merged/test/map \
--result_dir results/MnihCNN_multi_2016-02-03_03-34-58/ma_prediction_400 \
--channel 3 \
--offset 8 \
--relax 3 \
--steps 1024

Results

Conventional methods

Multi-channel models (epoch = 400, step = 1024)

Test on urban areas (epoch = 400, step = 1024)

x0_sigma for inverting feature maps

159.348674296

After prediction for single MA

$ bash shells/predict.sh
$ python scripts/integrate.py --result_dir results --epoch 200 --size 7,60
$ PYTHONPATH=".":$PYTHONPATH python scripts/evaluate.py --map_dir data/mass_merged/test/map --result_dir results/integrated_200 --channel 3 --offset 8 --relax 3 --steps 256
$ PYTHONPATH="." python scripts/eval_urban.py --result_dir results/integrated_200 --test_map_dir data/mass_merged/test/map --steps 256

Pre-trained models and Predicted results

• Pre-trained models
• Predicted results

Reference

If you use this code for your project, please cite this journal paper:

• Multiple Object Extraction from Aerial Imagery with Convolutional Neural Networks (bibtex)

Shunta Saito, Takayoshi Yamashita, Yoshimitsu Aoki, "Multiple Object Extraction from Aerial Imagery with Convolutional Neural Networks", Journal of Imaging Science and Technology, Vol. 60, No. 1, pp. 10402-1-10402-9, 2015