DeepTAM
DeepTAM is a learnt system for keyframe-based dense camera tracking and mapping.
If you use this code for research, please cite the following paper:
@InProceedings{ZUB18,
author = "H. Zhou and B. Ummenhofer and T. Brox",
title = "DeepTAM: Deep Tracking and Mapping",
booktitle = "European Conference on Computer Vision (ECCV)",
month = " ",
year = "2018",
url = "http://lmb.informatik.uni-freiburg.de/Publications/2018/ZUB18"
}
See the project page for the paper and other material.
Note: Currently we only provide deployment code for the camera tracking. The mapping code will come soon.
Setup
Current version is tested on Ubuntu 16.04 and with Python3.
# install virtualenv manager (here we use pew)
pip3 install pew
# create virtualenv
pew new deeptam
# switch to virtualenv
pew in deeptam
# install tensorflow 1.4.0 with gpu
pip3 install tensorflow-gpu==1.4.0
# install some python modules
pip3 install minieigen
pip3 install scikit-image
# clone and build lmbspecialops (use branch deeptam)
git clone -b deeptam https://github.com/lmb-freiburg/lmbspecialops.git
LMBSPECIALOPS_DIR=$PWD/lmbspecialops
cd $LMBSPECIALOPS_DIR
mkdir build
cd build
cmake ..
make
# add lmbspecialops to your PYTHON_PATH
pew add $LMBSPECIALOPS_DIR/python
# clone deeptam git (currently only tracking code is available)
git clone https://github.com/lmb-freiburg/deeptam.git
DEEPTAM_DIR=$PWD/deeptam
# add deeptam_tracker to your PYTHON_PATH
pew add $DEEPTAM_DIR/tracking/python
Running tracking examples
# download example data
cd $DEEPTAM_DIR/tracking/data
./download_testdata.sh
# download weights
cd $DEEPTAM_DIR/tracking/weights
./download_weights.sh
The basic example shows how to use DeepTAM to track the camera within one keyframe:
# run a basic example
cd $DEEPTAM_DIR/tracking/examples
python3 example_basic.py
The advanced example shows how to track a video sequence with multiple keyframes:
# run an advanced example
cd $DEEPTAM_DIR/tracking/examples
python3 example_advanced_sequence.py
# or run without visualization for speedup
python3 example_advanced_sequence.py --disable_vis
License
deeptam is under the GNU General Public License v3.0