EuRoC Forum

[michieletto]

Dear all,
trying out our code for the Task 1.1 we found that are some strange behaviors in the ground truth for a number of example trajectories.
- Sometimes the human swings and subsequently falls down.
- The trajectory index and the point cloud are not well aligned when published and no timestamps are provided to identify the exact beginning of a new trajectory.
- The ground truth returns true negatives. I.e. the human goes in the opposite direction with respect to the object and its point cloud it's not even visible, while the ground truth says he's pointing.

In the video you can find in a small test the two latter behaviors and the oscillation of the former one .

Code: Select all

http://robotics.dei.unipd.it/images/EuRoC/euroc_track1_1.mp4

Has anyone notice similar behaviors?

Thank you.
Best.
Stefano

[ipa-lth]

Dear Stefano,
we are working on some bugfixes and hopefully these fixes will solve your issues too.

Besides the "falling over" of the human model (we are working on that), I was not able to notice any similar behaviours so far.

If you have more information please let me know.

Best,
Lorenz

[michieletto]

Dear Lorenz,
thank you for your answer!

Regarding the human model behavior, actually, we are not able to find a pattern behind that and we noticed it only three times so far. We will inform you if we find out something more.

Best.
Stefano

[ipa-lth]

Dear Stefano,
the updates according the human model should be available now.

Unfortunately, I was neither able to reproduce your bug, nor able to fully understand what was happening. Furthermore it seems that the video is offline now.

Maybe the following information help you:
- The value of the ground truth data is set right from the moment the trajectory starts. It indicates that at some point during the following trajectory the human will point or not
- Very short time after a new gt data is set, the trajectory index changes
- One second after the trajectory index the new human trajectory is triggered
- The time delays between the sending of each signal are on purpose, to give a very high possibility, that the asynchronous ros topic communication has happened in the above order.

If the problems persists after the update,
please check for the computational load of your pc or support us with some more information.

Best,
Lorenz

[altella]

Hello all;
From my side, I think there is still some problems with the groundtruth of task 1.1.
There are still indications of "true gestures" when there is no gesture.
Besides, in simulation, it is intented to have 10 true positives and 10 true negatives. In my last tests, I am only having 8 gestures out of 20.

Kind regards,

Alberto

[ipa-lth]

Hi Alberto,
The mixture of human trajectories is random in the training version.

In fact I have never seen such a case.
Does the error occure if you do not visualize the point-clouds in rviz?
Do you reach the limits of the processor you are using? Are you loosing timestamps? What is the frame rate in rviz and or gazebo?
Can you check for the topic rate of the pointcloud msgs?
Can you provide your log files of such an misbehaviour case (ideally in DEBUG mode). Can you try to observer what the trajectory name is in the info log while failing.

Thanks for your help,
Lorenz

[altella]

Hi Lorenz;

Thanks a lot for your quick response !
The trajectories are random, so one thing clarified.
I try to check the others.

Thanks again,

Alberto

[admin]

Hi Alberto,
regarding your question:

Besides, in simulation, it is intented to have 10 true positives and 10 true negatives. In my last tests, I am only having 8 gestures out of 20.

We found a bug causing the number of true gestures be different from 10.

We will provide an update on the simulator that fixes this issue in ca. 1 hour.

Best