trajectory in a loop

[luigifreda]

trajectory in a loop

[luigifreda]

I would start considering the following points:

• Did you correctly calibrate your camera?
• If you use the camera behind a windshield, the windshield itself may act as an additional 'distorting' lens, therefore, you need further calibration for that.
• Does your camera have autofocus? If yes, that can change your calibration parameters and continuously break the assumption of a constant calibration matrix.

In general, in order to have a better understanding of what's going on, one needs to take a look at the video and consider all the parameter settings (if you changed them), the used features, etc.

[feritte]

Thanks for your answer. I'll definetely verify calibration and redo if it is necessary considering your remarks. What is the color code (Blue vs green path) in the mapping window?

[luigifreda]

• covisibility graph in green
• spanning tree in blue
  take a look at ORBSLAM paper

[feritte]

thanks a lot for your answer. I am reading the paper for refreshing the theory. Is it also possible that the warnings that I am receiving at the beginning of the main_slam.py can cause this performance issue ?

-WARNING: cannot import ORBextractor from orbslam2_features, check the file TROUBLESHOOTING.md

-WARNING: cannot import pyslam_utils

-WARNING: cannot import Orbslam2Feature2D from feature_orbslam2, check the file TROUBLESHOOTING.md

[luigifreda]

Those warnings mean that you didn't correctly compile C++ libraries (wrapped with pybind11) or your working python environment does not succeed to correctly import them. You may want to test the new install_all_conda.sh script.
I don't know which local features you are actually using. Consider that there are local features in the list that are not really working well as expected.

[feritte]

well I run the new script and all the installation thirdparty included was successful but I have import g2o
ModuleNotFoundError: No module named 'g2o' error when I tried to run main_slam.py

[feritte]

deleting build ../g2o/build and repeating the g2o installation fixed that error.. however the same warnings are still there.

[luigifreda]

You can use the script clean.sh in order to clean things in a proper way. Then, you have to delete your pyslam environment with pyenv-conda-delete.sh and launch install_all_conda.sh again.

[feritte]

i followed the steps. but still i have the same warnings therefore when I choose orb2 feature FeatureTrackerConfigs.ORB2 ==> self._feature_detector = Orbslam2Feature2D(self.num_features, self.scale_factor, orb2_num_levels)
TypeError: 'NoneType' object is not callable

[luigifreda]

Same problems here. Worked again on that. I've just updated the install scripts. You can now retry and everything should work.

[feritte]

thanks a lot Luigi. latest update has fixed the problems with binding. I dont have warnings anymore...:)
i will try my small video sequence again with my new calibration file. Just to be sure, for example videos from kittim there is always groundtruth.txt file and even main_slam.py also reads it but it needs this file in order to have good results ?
thanks again

[luigifreda]

I'm glad it worked.
For the sake of clarity, I have set groundtruth=None in main_slam.py. The ground-truth is not used at all by the Slam() class. You may want to better analyze the code and study the suggested references before actually keeping on with further experiments :-)