TreeSpotter

Spots street trees!

As requested by this poster on the C line of the NYC subway:

treescount.nycgovparks.org

Development

Please fork the repo and submit pull requests as needed.

The data directory holds scripts for pulling down datasets. The directory also stores the actual datasets when fetched, but note that these should not be committed into the repo. Use .gitignore files to avoid such issues.

The models directory holds Caffe model descriptions. It may also store trained weights which should not be committed into the repo. Again, use .gitignore files to avoid issues.

The tools directory contains subproject code in its subdirectories. Any new code should exist in a subdirectory of the tools directory.

Please follow these coding style guides where relevant:

• Python
• C++
• Shell
• JavaScript

Data Setup

Please request the URLs to our cached maps. Then paste them into the following config files:

echo "PRIVATELY_SHARED_URL/2012.tar.gz" > data/config/2012_MAP_URL.txt
echo "PRIVATELY_SHARED_URL/2014.tar.gz" > data/config/2014_MAP_URL.txt

Now let's fetch all raw data:

cd data

# Fetch NYC street tree census
make fetch_trees

# Fetch NYC satellite maps
make fetch_cached_maps

# Fetch NYC street data
make fetch_streets

cd ..

Now let's munge the data for training:

# Install Python script dependencies
# (alternatively use virtualenv if you don't want system level install)
sudo pip install -r tools/street_finder/requirements.txt
sudo pip install -r tools/cut_training_images/requirements.txt

## You may need add this to your env on OSX:
# export DYLD_FALLBACK_LIBRARY_PATH=$DYLD_FALLBACK_LIBRARY_PATH:$(HOME)/lib:/usr/local/lib:/lib:/usr/lib

# Create a CSV of streets from the source KML
python tools/street_finder --streets_dir data/streets

# Slice up positive and negative examples to train and test on
python tools/cut_training_images --trees_dir data/trees --map_dir data/maps/2014 --streets_dir data/streets --slices_dir data/slices/2014_64px --slice_size=64

# Create training list of images for caffe
./data/scripts/create_image_data_list.sh data/slices/2014_64px 0.7 > data/slices/2014_64px/train.txt
./data/scripts/shuffle_lines.sh data/slices/2014_64px/train.txt > data/slices/2014_64px/train_shuffled.txt

# Create testing list of images for caffe
./data/scripts/create_image_data_list.sh data/slices/2014_64px -0.7 > data/slices/2014_64px/test.txt
./data/scripts/shuffle_lines.sh data/slices/2014_64px/test.txt > data/slices/2014_64px/test_shuffled.txt

Training the Naive Model

A naive three-layer model is included as naive_64px

To train this model:

./models/scripts/train.sh naive_64px

After training, you can test it against the test data:

./models/scripts/test.sh naive_64px

Training a Custom Model

If you don't want to start from scratch, simply copy the naive model:

cp -rf models/naive_64px models/my_net

Now edit:

• the name value in models/my_net/train_test.prototxt
• the name value in models/my_net/deploy.prototxt
• the net and snapshot_prefix paths in models/my_net/solver.prototxt

You can now tweak the architecture of the network in models/my_net/train_test.prototxt and solver hyperparamters in models/my_net/solver.prototxt.

To train your network:

./models/scripts/train.sh my_net

To test your network: ./models/scripts/test.sh my_net