Run ORB-SLAM3

What you will do

You will build ORB-SLAM3 following its README, run three sensor configurations on one EuRoC sequence, and evaluate the trajectories. Build details change with the commit and your operating system: follow the README and its issue tracker for the version you use; the steps below give the structure and the checks. The code is GPLv3.

Before you start

• A Linux machine with the dependencies named in the ORB-SLAM3 README (Pangolin, OpenCV, Eigen3, a C++ compiler), about 10 GB of disk and the EuRoC MAV dataset sequence MH_01_easy (ASL format) from the ETH dataset page.
• Lessons M4-T2, M4-T3 and the lab “Analyze trajectory drift”.

Steps

1. Clone the repository, read its README and install the listed dependencies; build with the provided build.sh.
2. Download MH_01_easy in ASL format and note the path.
3. Run monocular, stereo and stereo-inertial EuRoC examples (executable names and arguments are in the README); keep the output trajectory file of each.
4. Convert the trajectories to TUM format if needed (the ORB-SLAM3 outputs are in TUM-like format for the keyframe and full trajectories), then evaluate each with evo_ape tum gt.txt est.txt -va --align (add --correct_scale for monocular).
5. Record ATE, whether the system lost tracking (watch the viewer output) and the average frame processing time.

# structure of the commands (check the README of your commit for exact paths and arguments)
./Examples/Monocular/mono_euroc ./Vocabulary/ORBvoc.txt ./Examples/Monocular/EuRoC.yaml PATH/MH01 ./Examples/Monocular/EuRoC_TimeStamps/MH01.txt dataset-MH01_mono
./Examples/Stereo-Inertial/stereo_inertial_euroc ./Vocabulary/ORBvoc.txt ./Examples/Stereo-Inertial/EuRoC.yaml PATH/MH01 ./Examples/Stereo-Inertial/EuRoC_TimeStamps/MH01.txt dataset-MH01_stereoi
evo_ape tum MH01_gt_tum.txt f_dataset-MH01_stereoi.txt -va --align --plot --plot_mode xy

Experiments

• Disable the IMU on the stereo run (use the stereo example) and compare drift.
• Halve the image resolution and the feature count in the settings file and record accuracy and speed.
• Run on a harder sequence (MH_04_difficult) and note where tracking is lost.

How to know it worked

• Three trajectories, each with an ATE number and the alignment option stated.
• A table of ATE and processing time per configuration.
• A statement of where tracking failed and why, from the viewer and the feature counts.

If it does not work

References

• ORB-SLAM3 repository and README
• EuRoC MAV dataset (ETH ASL)
• evo

Record your progress