This repository provides software for our publication "End-to-End Training of Hybrid CNN-CRF Models for Stereo", which is going to be presented at CVPR 2017.

If you use this code please cite the following publication:

@inproceedings{knoebelreiter_cvpr2017,
  author = {Patrick Knöbelreiter and Christian Reinbacher and Alexander Shekhovtsov and Thomas Pock},
  title = {{End-to-End Training of Hybrid CNN-CRF Models for Stereo}},
  booktitle = {2017 Computer Vision and Pattern Recognition (CVPR)},
  year = {2017},
}

The repository is structured as follows:

• stereo contains the C++/CUDA code of our CNN-CRF model -> this needs to be compiled to use our model for disparity computation
• undistort contains python code for rectifying the Middlebury images
• eval contains python code for reproducing the numbers presented in the paper
• data contains the learned model parameters

For your convenience, the required libraries that are on Github are added as submodules. So clone this repository with --recursive or do a

git submodule update --init --recursive

after cloning. All dependencies will be located in the dependency folder like

dependencies/cnpy
dependencies/imageutilities
dependencies/slackprop

This software requies:

• GCC >= 4.9
• CMake >= 3.2
• ImageUtilities (https://github.com/VLOGroup/imageutilities) with the iuio and iumath modules
• SlackProp (https://github.com/VLOGroup/slackprop)
• cnpy (https://github.com/rogersce/cnpy)
• cudnn 6.x (https://developer.nvidia.com/cudnn)

Compile and install imageutilities: Follow the instructions on https://github.com/VLOGroup/imageutilities. Make sure you set the environment variable IMAGEUTILITIES_ROOT correctly. This is necessary to find the compiled library automatically with CMake.

Compile and install cnpy by executing the following commands:

cd dependencies/cnpy
mkdir build
cd build
cmake ..
(sudo) make install
cd ../../

More information can be found at https://github.com/rogersce/cnpy.

Compile SlackProp using the following commands:

cd dependencies/slackprop
mkdir build
cd build
cmake ..
make 
cd ../../

At this point you should have compiled all the dependencies successfully. Please also double-check you have set the environment variable IMAGEUTILITIES_ROOT correctly.

Compiling our stereo model:

mkdir build
cd build
cmake ../stereo
make

If everything worked correctly, you should see an executeable called stereo_img in your build directory. The following simple test will print the usage information

./stereo_img

In order to demonstrate the usage of our code we put a rectified stereo-pair into the data directory. You can compute the disparity map using

./stereo_img --im0 ../data/im0.png --im1 ../data/im1.png

This will create a file called output.png in the same directory. Otherwise, download the Middlebury data as described below. Then you can test the algorithm using

./stereo_img --im0 ../data/middlebury-2014/MiddEval3/trainingQ/Adirondack/im0.png  --im1 ../data/middlebury-2014/MiddEval3/trainingQ/Adirondack/im1.png --parameter-file ../data/parameters/middlebury-2014/7-layer/cnn+crf+full/params --config-file ../data/parameters/middlebury-2014/7-layer/cnn+crf+full/config_mb_cnn7_crf_full.cfg

• OpenCV >= 3.0 (https://github.com/opencv/opencv) with python bindings
• OpenCV contribution package xfeatures2d (https://github.com/opencv/opencv_contrib) with python bindings
• scipy >= 0.17
• The C-Shell csh

First you must download the data from the respective benchmark and then you can use the provided evaluation scripts to reproduce the numbers in the paper.

Create the two directories middlebury-2014 and kitti-2015 in the data directory, such that the structure of the data directory is

data/kitti-2015/
data/middlebury-2014/

1. Download the Middlebury 2014 data from (http://vision.middlebury.edu/stereo/submit3/). You will will need Input Data as well as Ground truth for left view for quarter (Q), half (H) and full (F) resolution.
2. Download the Middlebury evaluation SDK (http://vision.middlebury.edu/stereo/submit3/zip/MiddEval3-SDK-1.6.zip)
3. Extract all downloaded files to the data folder. Your data folder should look like

data/
|--- middlebury-2014/
    |--- MiddEval3/
        |--- trainingQ/
        |--- trainingH/
        |--- testQ/
        |--- testH/
        |--- runevalF
        |--- ...

4. Compile the Middlebury evaluation SDK like described in http://vision.middlebury.edu/stereo/submit3/zip/MiddEval3/README.txt
5. Rectify train/test images using the provided script

cd undistort
python main.py ../data/middlebury-2014/MiddEval3/trainingH/
python main.py ../data/middlebury-2014/MiddEval3/testH/

This command will warp im1 such that corresponding pixels are located in the same row. The rectified images are saved as im1_rectified.png in the appropriate folder.

6. Compute results for Middlebury

cd eval
./run_all_middlebury.sh

7. Compute Errors

python compute_numbers_middlebury.py

1. Download the Kitti 2015 data from (http://www.cvlibs.net/download.php?file=data_scene_flow.zip).
2. Make a folder kitti-2015 in the data folder and extract all downloaded files there Your data folder should look like

data/
|--- kitti-2015
    |--- testing/
    |--- training/

3. Compute results for Kitti

cd eval
./run_all_kitti_2015.sh

4. Compute Errors

python compute_numbers_kitti.py