fcn 4.0.0

Fully Convolutional Networks

fcn - Fully Convolutional Networks
==================================

.. image:: https://badge.fury.io/gh/wkentaro%2Ffcn.svg
:target: https://badge.fury.io/gh/wkentaro%2Ffcn
.. image:: https://travis-ci.org/wkentaro/fcn.svg?branch=master
:target: https://travis-ci.org/wkentaro/fcn


This is Chainer_ implementation of fcn.berkeleyvision.org_.

.. _fcn.berkeleyvision.org: https://github.com/shelhamer/fcn.berkeleyvision.org.git
.. _Chainer: https://github.com/pfnet/chainer.git


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_]

.. _v1.0.0: https://github.com/wkentaro/fcn/releases/tag/v1.0.0
.. _v2.0.0: https://github.com/wkentaro/fcn/releases/tag/v2.0.0
.. _v3.0.0: https://github.com/wkentaro/fcn/releases/tag/v3.0.0


License
-------
| Copyright (C) 2016 Kentaro Wada
| Released under the MIT license
| http://opensource.org/licenses/mit-license.php


For Beginners
-------------

Installation
++++++++++++

.. code-block:: bash

# LINUX: install required libraries via apt
sudo apt-get install liblapack-dev # for numpy
sudo apt-get install libhdf5-dev # for h5py
sudo apt-get install libleveldb-dev # for plyvel
# OS X: install required libraries via brew
brew install gfortran # for numpy
brew install leveldb # for plyvel
brew install hdf5 # for h5py

pip install -U setuptools
pip install fcn

fcn_install_trained_data.py


Forwarding
++++++++++

Forwarding is done as below, and computation graph is `here <https://github.com/wkentaro/fcn/blob/master/_images/fcn8s_forward.jpg>`_.

.. code-block:: bash

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

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

.. image:: https://raw.githubusercontent.com/wkentaro/fcn/master/_images/fcn8s_26471792680.jpg

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


For Developers
--------------


Installation
++++++++++++

.. code-block:: bash

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

python setup.py install
fcn_install_trained_data.py

You need to download pascal VOC2012 dataset from `here <http://host.robots.ox.ac.uk/pascal/VOC/voc2012/>`_,
and install it as below construction::

- fcn - data - pascal - VOC2012 -- JPEGImages
- SegmentationClass
- ...

Fowarding
+++++++++

Forwarding with trained model in caffe is done as below:

.. code-block:: bash

# This downloads caffemodel and convert it to chainermodel
./scripts/caffe_to_chainermodel.py

# forwarding of the networks
./scripts/fcn_forward.py --img-files data/pascal/VOC2012/JPEGImages/2007_000129.jpg

.. image:: https://raw.githubusercontent.com/wkentaro/fcn/master/_images/fcn8s_2007_000129.jpg

Original Image: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/


Training
++++++++

.. code-block:: bash

./scripts/fcn_train.py

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

.. image:: https://raw.githubusercontent.com/wkentaro/fcn/master/_images/fcn32s_learning_curve.png

Forwarding with ``fcn32s_60000.chainermodel`` ends with below result:

.. image:: https://raw.githubusercontent.com/wkentaro/fcn/master/_images/fcn32s_2007_000129.jpg