Data-Efficient Decentralized Visual SLAM

This is the code for the 2018 ICRA paper Data-Efficient Decentralized Visual SLAM by Titus Cieslewski, Siddharth Choudhary and Davide Scaramuzza:

@Article{Cieslewski18icra,
  author        = {Titus Cieslewski and Siddharth Choudhary and Davide
                  Scaramuzza},
  title         = {Data-Efficient Decentralized Visual {SLAM}},
  journal       = {{IEEE} Int. Conf. Robot. Autom. ({ICRA})},
  year          = 2018,
  month         = may
}

You can find a pdf of the paper here. If you use any of this code, please cite this publication.

You might be looking for the python/tensorflow implementation of NetVLAD, which is also tied to that publication. Find that one here.

Downloads

File	Size	Description	sha1
kitti.zip	144.7MB	Intermediate files for KITTI 00	f5f39202afdb598d50dc14d43580114cd7eb81bf
stata.zip	649.8MB	Intermediate files for Stata 2012-01-25-12-14-25	8b6c488ce3ca99b7abd41a553387ef9c7c447a24
robotcar_netvlad_feats.zip	414MB	NetVLAD features used for DVPR cluster training	8785d63dd56b15ebc8de8a6b1c7e3554754953e4

Install instructions

C++

Install the ROS ecosystem. This has been developed with ROS version Jade under Ubuntu 14.04. Minimal installation will do. Next, install catkin tools, vcstool, OpenCV-nonfree dev, autoconf and libglew-dev:

sudo add-apt-repository --yes ppa:xqms/opencv-nonfree # Maybe not necessary for Ubuntu > 14.04?
sudo apt-get update
sudo apt-get install python-catkin-tools python-vcstool libopencv-nonfree-dev autoconf libglew-dev

# Create a new catkin workspace if needed:
mkdir -p my_ws/src
cd my_ws
catkin config --init --mkdirs --extend /opt/ros/<YOUR VERSION> --merge-devel --cmake-args -DCMAKE_BUILD_TYPE=Release

# Clone dslam:
cd src
git clone git@github.com:uzh-rpg/dslam_open.git

# Clone dependencies:
vcs-import < dslam_open/dependencies.yaml

# Build:
catkin build

Matlab

For full reproduction, you need to install NetVLAD and all its dependencies and add it to your matlab path. However, for the datasets evaluated in the paper, we provide the intermediate results generated by NetVLAD, so you can skip the installation of NetVLAD for now.

Distributed Trajectory Estimation

Build distributed-mapper in the feature/logging branch in an arbitrary location. No need to build the unit tests (they require extra dependencies), all you need is the runDistributedMapper executable.

Run instructions

In the following, we list all data processing steps that produce intermediate results. All intermediate results are saved in <DATASET SEQUENCE ROOT>/dslam/. Furthermore, we provide all intermediate results for the datasets evaluated in the paper, so you can skip the corresponding steps and just use our intermediate results. Note that some steps are nondeterministic (ORB SLAM, NetVLAD clustering, the decentralized SLAM simulation itself).

Dataset preprocessing

Note that these are only necessary if you want to reproduce the pipeline fully or want to adapt the pipeline to your own dataset. If you just want to run the system on either of the two datasets in the paper, you can use the preprocessed data we provide and skip straight to Execute the remaining preprocessing sections.

KITTI 00

For KITTI 00, obtained here, no special preprocessing is necessary.

Stata 2012-01-25-12-14-25

For Stata 2012-01-25-12-14-25, which can be obtained here, we require somewhat more preprocessing. We need to first rectify the images and then parse images, times and ground truth poses out of the rosbag. Launch dslam/launch/rectify_stata.launch, start recording /wide_stereo/left/image_rect and /wide_stereo/right/image_rect into a rosbag, and play back the main dataset rosbag. Once the images are rectified, run dslam/src/stata_unbag.py with the following parameters:

• --bag_name path to the main bag downloaded from the dataset website
• --rect_bag_name path to the bag with rectified images you just created
• --outdir output directory
• --tf_file the ground truth .poses file downloaded from the stata dataset (rightmost column in the downloads table)

This will take some time and create the images folder in the output directory, as well as times.txt, gt_poses.txt and some auxiliary files that help you assemble the ORB SLAM settings file (or you can use the one we provide).

Run ORB SLAM (loop closure disabled) to generate visual odometry data

In your catkin workspace, run dataset_to_vo, for example on KITTI 00:

./devel/lib/dslam/dataset_to_vo --type=kitti /home/titus/data/kitti/00/ /home/titus/data/kitti/00/orb_kitti.yaml

Where orb_kitti.yaml are the configuration parameters to run ORB SLAM on the KITTI datasets. They are available in the files we provide for download. Once this has completed, the following files will be created in the current working directory:

descs_0.txt
lm_obs_0.txt
lm_pos_0.txt
poses_0.txt
times_0.txt

Copy these files to <DATASET SEQUENCE ROOT>/dslam/. If more than one set of these exists, ORB SLAM failed at some point. In that case, either rerun the simulation until it does not fail, or, if you have one set containing the majority of the dataset, just use that set (rename it to end with _0.txt).

Configure the main simulation parameters in decentr_sim.m

All subsequent steps will take place in the Matlab script dslam/matlab/decentr_sim.m, so open that script in Matlab and set the configuration variables as indicated in the first section. Here is an example configuration:

Note that dvpr_train_path is the path containing NetVLAD features to be used for the clustering of the NetVLAD space used by decentralized visual place recognition. See the corresponding paper for detailed insight. As features, we used a random bunch of features extracted from several sequences of the Oxford RobotCar dataset. These features are available as download and we recommend you use them as well. Alternatively, you can have the variable point to any folder containing several .bin files with NetVLAD output, representing your NetVLAD clustering training dataset.

Parse NetVLAD descriptors for the dataset

Note: This is done with the Matlab implementation of NetVLAD for dslam. If you are looking for the python/tensorflow implementation of NetVLAD that is also tied to this publication, you will find it here.

This step takes quite some time (10 minutes on KITTI 00 with an NVIDIA GT 1060), so consider using the provided file instead. If you still want to run this yourself, in decentr_sim.m, execute the section %% Generate NetVLAD vectors for dataset. Note that this requires NetVLAD weights. We used vd16_pitts30k_conv5_3_vlad_preL2_intra_white.mat provided here. Once the parsing is complete, the file netvlad_feats.bin will appear in <DATASET SEQUENCE ROOT>/dslam/.

Note that NetVLAD can only take jpg pictures as input. You might need to convert any png pictures into jpg, for example using the following bash commands:

for i in *.png ; do convert "$i" "${i%.*}.jpg" ; done

Execute the remaining preprocessing sections

The next three sections in decentr_sim.m take you through the rest of preprocessing that sets up the decentralized SLAM simulation:

1. parseAllData() zips up the visual odometry (VO) output from ORB SLAM and the NetVLAD descriptors parsed in the previous section and puts the result into full_data.mat.
2. splitAndSave() subdivides the VO + NetVLAD data into n subsequences, one per simulated robot. The result is saved in full_data_n_robots.mat, where n is replaced with the used number of robots.
3. generateDecentrFrameStream() takes the subsequences and rearranges the VO data into a stream representing the VO of all sequences being played back simultaneously. The stream is saved into decentr_stream_n_robots.mat.

Simulate!

Now that all preprocessing is done, you are ready to run the decentralized SLAM simulation!

1. Tweak the parameters in the %% Parameters section.
2. Don't forget to have one process of the verification_request_server running.

If you are running inside the dslam/matlab directory, the verification_request_server can be run with:

 ../../../../devel/lib/dslam/verification_request_server temp_request.txt temp_result.txt temp_lock.txt

Execute up until and with the runSimulationAndSave() command. This command will run the simulation and give you live visual feedback like this (see also gif on top of readme):

Once the simulation is complete, results will be written into a file called run_data_X.mat, where X contains a string of the parameters used in the simulation.

Plot

Once the simulation is complete, you can re-plot the paper figures for your run with the sections following runSimulationAndSave() in decentr_sim.m. If you want to reproduce the parameter studies (min dist gv parameter study and NetVLAD dimension parameter study), you can find the code for them in param_studies.m.

Acknowledgements

This research was funded by the DARPA FLA Program, the National Center of Competence in Research (NCCR) Robotics through the Swiss National Science Foundation, the SNSF-ERC Starting Grant and the U.S. Army Research Lab (ARL) through the MAST-CTA project 329420