EurocMono params on master gives poor VIO performance

[sarika93]

Description:
Running the master branch using the stereoVIOeuroc.bash script with Euroc V1_01_easy dataset using the default EurocMono params on master gives very poor VIO performance. The pipeline runs until completion. Note: stereo performance is still good at less than +/-0.1m of error in each axis with the same dataset.

The below shows the error performance with the default EurocMono setup:

This was not the case in an older version of master (git hash: a2cf0af), which gave the following error performance.

It looks like with OpenCV or GTSAM related updates something was broken.

However, I was able to find a working config which gives good VIO performance when running with master. Changing the frontend detector from GFTT (feature_detector_type: 3 in FrontendParams.yaml) to FAST (feature_detector_type: 0 in FrontendParams.yaml) , the performance improves signficantly. See below:

Changing the linearization mode from Hessian to Implicit Schur (linearizationMode: 1 in BackendParams.yaml) further improves performance:

All graphs were generated using this script.

I will be submitting a PR with these updated params for the EurocMono configuration. However, this doesn't fix any possible underlying issues here.

Steps to reproduce error:

1. Checkout this repository to the latest version of master
2. Update the scripts/stereoVIOeuroc.bash file to use EurocMono params and point to the relevant dataset path. Output should also be logged.
3. Run scripts/stereoVIOeuroc.bash. Using the visualizer, the explosion in pose estimation should be easily visible. Alternatively, the script linked above can be used to produce an error graph.

My testing setup

• Kimera-Vio branch, tag or commit used: master, git hash: acf45cb
• GTSAM version used: 4.2
• OpenGV version used:
• OpenCV version used: 4.2.0
• Operating system and version : Ubuntu 20.04 via Docker
• Did you change the source code? (yes / no): No, this was tested on the master branch.