This package allows researches to apply the described networks to match image patches and extract corresponding patches.

We tried to make the code as easy to use as possible. The original models were trained with Torch ( http:https://torch.ch ) and we release the models in Torch7 format, and in addition to that we have C++ bindings which do not require Torch installation. Thus we provide example code how to use the models in MATLAB and with OpenCV http:https://opencv.org

Copyright © 2015 Ecole des Ponts, Universite Paris-Est

All Rights Reserved. A license to use and copy this software and its documentation solely for your internal research and evaluation purposes, without fee and without a signed licensing agreement, is hereby granted upon your download of the software, through which you agree to the following:

1. the above copyright notice, this paragraph and the following three paragraphs will prominently appear in all internal copies and modifications;
2. no rights to sublicense or further distribute this software are granted;
3. no rights to modify this software are granted; and
4. no rights to assign this license are granted.

Please Contact Prof. Nikos Komodakis, 6 Avenue Blaise Pascal - Cite Descartes, Champs-sur-Marne, 77455 Marne-la-Vallee cedex 2, France, Office B003 Tel : +33164152173, Fax: +33164152186 for commercial licensing opportunities, or for further distribution, modification or license rights.

Created by Sergey Zagoruyko and Nikos Komodakis. http:https://imagine.enpc.fr/~komodakn/

Please cite the paper below if you use this code in your research.

Sergey Zagoruyko, Nikos Komodakis "Learning to Compare Image Patches via Convolutional Neural Networks", ArXiv:1504.03641. http:https://arxiv.org/abs/1504.03641

The code was tested to work in Linux (Ubuntu 14.04) and OS X 10.10, although we release all the source code to enable usage in other operating systems.

We provide the models in 3 formats, two are Torch7 "nn" and "cudnn" formats and one is binary format with weights only. The table from the paper is here for convenience.

To save time downloading models one by one run the script (downloads ~375MB):

./download_pack.sh

Models in nn format can be loaded and used without CUDA support in Torch. To enable CUDA support model:cuda() call required.

Models in cudnn format are faster, but need a special library from NVIDIA. Check https://github.com/soumith/cudnn.torch

We release CUDA code for the time being, CPU code might be added in the future. To install it you need to have CUDA 7.0 with the up-to-date CUDA driver. Clone and compile this repository it with:

git clone --recursive https://github.com/szagoruyko/cvpr15deepmatch
cd cvpr15deepmatch
mkdir build; cd build;
cmake .. -DCMAKE_INSTALL_PREFIX=../install
make -j4 install

OpenCV example is here to demonstrate how to use the deep CNN models to match image patches, how to preprocess the patches and use the proposed API.