This repository contains a working Docker implementation of ORB-SLAM3. This implementation is heavily based on Kevin Robb's.
Original ORB-SLAM3 repo: https://github.com/UZ-SLAMLab/ORB_SLAM3 (last updated: Dec 2021)
Check orbslam3_explained for useful details on ORB-SLAM3.
For a exhaustive step-by-step log of how I implemented ORB-SLAM3 and a list of what still remains to be implemented, check HOW.
Clone this repository
git clone https://github.com/drodriguezSRL/ORBSLAM3_implementation
Run the dataset setup script. This downloads, unzips, and fixes the EuRoC dataset (some images may be corrupted and cause problems down the line) and places it in ./data/EuRoc.
cd ORBSLAM3_implementation
bash download_euroc_mh01.sh
Build and run the Dockerfile through docker-compose
cd docker
docker-compose build
docker-compose run orbslam3-spell
In its current version, the Docker container includes a cloned and fixed workspace but unbuilt version of ORB-SLAM3. We need to manually build it inside the container.
Within the container, run the following command:
cd ~/Dev/ORB_SLAM3
sh build.sh
Important
As indicated by Kevin Robb's implementation instructions, ORB-SLAM3 requires multiple build runs before it can be succesfully built without any errors or warnings. In my experience, it often required between 3 to 5 builds (i.e, running the sh build.sh command multiple times) before it was successfully built.
Once ORB-SLAM3 is built, we can run a few simulation examples from the EuRoc dataset previously downloaded.
Within the container, change to the ~/Dev/ORB_SLAM3/ directory, and choose one of the following to run. A map viewer as well as an image viewer should appear during the simulation. Once done, a f_dataset-MH01_...txt and a kf_dataset-MH01_...txt files are created within the ORB_SLAM3 folder containing the camera complete estimated trajectory and keyframe trajectory, respectively.
For a detail description of these ORB-SLAM3 commands, check orbslam3_explained.
# Mono
./Examples/Monocular/mono_euroc ./Vocabulary/ORBvoc.txt ./Examples/Monocular/EuRoC.yaml ~/Datasets/EuRoc/MH01 ./Examples/Monocular/EuRoC_TimeStamps/MH01.txt dataset-MH01_mono
# Mono + Inertial
./Examples/Monocular-Inertial/mono_inertial_euroc ./Vocabulary/ORBvoc.txt ./Examples/Monocular-Inertial/EuRoC.yaml ~/Datasets/EuRoc/MH01 ./Examples/Monocular-Inertial/EuRoC_TimeStamps/MH01.txt dataset-MH01_monoi
# Stereo
./Examples/Stereo/stereo_euroc ./Vocabulary/ORBvoc.txt ./Examples/Stereo/EuRoC.yaml ~/Datasets/EuRoc/MH01 ./Examples/Stereo/EuRoC_TimeStamps/MH01.txt dataset-MH01_stereo
# Stereo + Inertial
./Examples/Stereo-Inertial/stereo_inertial_euroc ./Vocabulary/ORBvoc.txt ./Examples/Stereo-Inertial/EuRoC.yaml ~/Datasets/EuRoc/MH01 ./Examples/Stereo-Inertial/EuRoC_TimeStamps/MH01.txt dataset-MH01_stereoi
The outcome can then be evaluated using the evaluate_ate_scale.py script as follows:
python2.7 tools/evaluate_ate_scale.py ./output/MH01_GT.txt ./output/f_dataset-MH01_mono.txt --verbose
Note
To use the evaluate_ate_scale.py script as is (i.e., as provided by ORB-SLAM3), you need python2.7 + numpy + matplotlib. You'll need pip2.7 for this:
sudo apt install curl
cd ~/Desktop
curl https://bootstrap.pypa.io/pip/2.7/get-pip.py --output get-pip.py
sudo python2 get-pip.py
pip2.7 install numpy matplotlib
The most straigthforward solution for making your own data work with ORB-SLAM3 would be to mimick as much as possible the directory layout and data structure of one of the working examples provided by the original ORB-SLAM3 implementation. Then, you'd simply need to call the pre-existing executable of your choosing (mono, mono-interial, stereo, or stereo-inertial) in a similar fashion as did before.
In my case, I chose to use the EuRoC data structure as a reference. Since making a one-size-fits-all guide for data adaption would be almost impossible (as it would depend on the nature of your own data to begin with), you can check here all the steps I followed to run ORB-SLAM3 on the SPICE-HL3 dataset.