DeMoN is "A computer algorithm for reconstructing a scene from two projections"¹. The network estimates the depth and relative camera motion for pairs of images; it addresses the important two view case in structure from motion.

If you use this code for research please cite:

@InProceedings{UZUMIDB17,
  author       = "B. Ummenhofer and H. Zhou and J. Uhrig and N. Mayer and E. Ilg and A. Dosovitskiy and T. Brox",
  title        = "DeMoN: Depth and Motion Network for Learning Monocular Stereo",
  booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
  month        = " ",
  year         = "2017",
  url          = "http://lmb.informatik.uni-freiburg.de//Publications/2017/UZUMIDB17"
}

See the project website for the paper and other material.

^{1 This is the title of H. C. Longuet-Higgins paper from 1981, which perfectly describes what our method does. DeMoN shows that complex geometric relations can be learnt by a ConvNet.}

Building and using requires the following libraries and programs

tensorflow 1.4.0
cmake 3.7.1
python 3.5
cuda 8.0.61 (required for gpu support)
VTK 7.1 with python3 interface (required for visualizing point clouds)

The versions match the configuration we have tested on an ubuntu 16.04 system. DeMoN can work with other versions of the aforementioned dependencies, e.g. tensorflow 1.3, but this is not well tested.

The binary package from vtk.org does not come with a python3 interface. To enable python3 support VTK needs to be built from source. Alternatively, there are also VTK packages with python3 support available in Anaconda via the conda package manager.

The network also depends on our lmbspecialops library which is included as a submodule.

The following describes how to install tensorflow and demon into a new virtualenv and run the inference example. We will use pew (pip3 install pew) to manage a new virtualenv named demon_venv in the following:

# create virtualenv
pew new demon_venv

The following commands all run inside the virtualenv:

# install python module dependencies
pip3 install tensorflow-gpu # or 'tensorflow' without gpu support
pip3 install pillow # for reading images
pip3 install matplotlib # required for visualizing depth maps
pip3 install Cython # required for visualizing point clouds

# clone repo with submodules
git clone --recursive https://github.com/lmb-freiburg/demon.git

# build lmbspecialops
DEMON_DIR=$PWD/demon
mkdir $DEMON_DIR/lmbspecialops/build
cd $DEMON_DIR/lmbspecialops/build
cmake .. # add '-DBUILD_WITH_CUDA=OFF' to build without gpu support
# (optional) run 'ccmake .' here to adjust settings for gpu code generation
make
pew add $DEMON_DIR/lmbspecialops/python # add to python path

# download weights
cd $DEMON_DIR/weights
./download_weights.sh

# run example
cd $DEMON_DIR/examples
python3 example.py # opens a window with the depth map (and the point cloud if vtk is available)

The data reader op and the evaluation code have additional dependencies. The code for the data reader is in the multivih5datareaderop directory. See the corresponding readme for more details.

For the evaluation see the example examples/evaluation.py. The evaluation code requires the following additional python3 packages, which can be installed with pip:

h5py
minieigen
pandas
scipy
scikit-image
xarray

Note that the evaluation code also depends on the data reader op.

Instructions for training a clean tensorflow version of DeMoN are here. Note that the tensorflow training code and model are work in progress and are not identical to the original Caffe version.

Download scripts for training and testing are located in the datasets subdirectory. Note that due to a bug that some of the dataset files with the prefix rgbd did contain some samples from the test set. The affected files have been replaced and now have the prefix rgbd_bugfix. MD5 checksums for all files can be found in the file traindata.md5.

DeMoN is under the GNU General Public License v3.0