fcn - Fully Convolutional Networks

This is Chainer implementation of fcn.berkeleyvision.org.

Features

• Provide FCN8s model for Chainer. [v1.0.0]
• Copy caffemodel to chainermodel. [v1.0.0]
• Forwarding with Chainer for pascal dataset. [v1.0.0]
• Training with Chainer for pascal dataset. [v2.0.0]
• Training for APC2015 dataset. [v3.0.0]

License

Copyright (C) 2016 Kentaro Wada

Released under the MIT license

http://opensource.org/licenses/mit-license.php

Installation

# Ubuntu: install required libraries via apt
sudo apt-get install liblapack-dev   # for numpy
sudo apt-get install libhdf5-dev     # for h5py
# macOS: install required libraries via brew
brew install gfortran  # for numpy
brew install hdf5      # for h5py

pip install fcn

For Inference

Inference is done as below:

# Download sample image
wget https://farm2.staticflickr.com/1522/26471792680_a485afb024_z_d.jpg -O sample.jpg

# forwaring of the networks
fcn_infer.py --img-files sample.jpg --gpu -1  # cpu mode
fcn_infer.py --img-files sample.jpg  # gpu mode

Original Image: https://www.flickr.com/photos/faceme/26471792680/

For Developing

git clone https://github.com/wkentaro/fcn.git
cd fcn

python setup.py develop

Training

cd examples/pascal
./download_dataset.py

./train_fcn32s.py

Currently we support only training FCN32s. The learning curve looks like below:

Inference with FCN32s + 60000 iterations outputs below result: