Using data.log as input

[jorou125]

Hi, thank you for making your code available. I am trying to run EDOPT on the data from your dataset but I am unable to do so. I am able to run the Docker and I have no errors from YARP. When i run the commands to run EDOPT on the dataset (for the tomato x 0.2m/s) i get a video of the car model stationary with no events (image included for context). Is it possible to explain how to use EDOPT on your dataset? Thank you

[arrenglover]

Hi @jorou125 - it looks like EDOPT can run successfully and load the model. If you want to use the tomato soup can dataset you will need to change the model that it is loading. try ./edopt --from ../configTOMATO.ini

secondly we want to run the data from the dataset and not the camera. You followed some instructions from here #13 correctly. There is only one thing I could imagine could cause failure if the yarpdataplayer is running and you successfully load the data, and press the play button. You should get something like this:

I'm using a different dataset in this example and you should have only 1 sensor (atis) loaded in the tomato soup dataset. in a terminal write yarp name list pay attention to the port names. There should be something like /atis/AE:o which should correspond to the port name written in the yarpdataplayer GUI. The other should be something like /edopt/AE:i. we need to connect the datastream to the algorithm using yarp connect /atis/AE:o /edopt/AE:i fast_tcp replacing the port names with the correct ones in your system in case I have an error here.

@lunagava has more recently run the code and might correct a mistake I made.

[jorou125]

Hi, thank you for the quick response! I may have found the source of my problems, but have yet found a solution. When I open yarpdataplayer there is nothing available. When i navigate to the folder containing the .log files, it seems I am able to access the data, but the port is not what it should be. Here is a screenshot for reference:

This tells me that I most likely did not initialize the docker correctly by mounting the data.log file. I used the following command:
docker run -it --privileged -v "/mnt/c/Users/jorou/OneDrive/Desktop/EDOPT/6-DoF_events/005_tomato/005_tomato_translation_x_0_2m_s:/data/005_tomato/005_tomato_translation_x_0_2m_s" -v /tmp/.X11-unix/:/tmp/.X11-unix -e DISPLAY=unix$DISPLAY --network host --gpus all --name edopt edopt:latest

Is this the appropriate way to run the docker and loading the data? Another possible cause is my use of the atis-sdk. When initializing it using the command to use a file and not a camera, it does not work. I get an error message that no camera is connected. Furthermore, I have tried ./edopt --from ../configTOMATO.ini, but get an error. It seems this is caused by the absence of the ini file in build. To correct this error I simply copied the .ini file. Now when I run the commands I do indeed have the tomato soup model, but still no events.

Lastly, I tried the yarp name list as suggested, but there is no /edopt/AE:i or /atis/AE:o. I assume this is linked to the issue with how I load the data.

Thank you for you support

[jorou125]

Hello @arrenglover, I was finally able to run EDOPT on the your dataset. I checked the code and the port isn't named /edopt/AE:i or /atis/AE:o. It is named /ekom/AE:i. Once I connect the port from the previous screenshot to this port I am able to visualize the events. I copied the config files in the build directory and now I am also able to see the correct model. Thank you for your help in sorting this out!

Although I can now see the events and the appropriate model, I am unable to reproduce the results from your article. Using the dragon model and the dragon_6DoF_1 data, I should be able to reproduce the results in your article but the tracking fails rather quickly (especially while using --parallel). Could you elaborate on how you produced your results for Fig.4 from the article so I can reproduce them. Thank you in advance!

[arrenglover]

Hi @jorou125 ,

Glad to hear that the pipeline seems to be up and running! Sorry that the tracking doesn't work immediately. I think this could be 2 issues.

1. parameter choice might have changed between running these particular experiments, and us running demos with different hardware and lighting conditions.
2. the algorithm is written to be "always on-line/real-time" so your laptop and GPU configuration can affect the speed at which the tracking update occurs, and therefore if failure occurs. We were typically hitting 200-300 Hz when we did the experiments for the paper. Some laptops it only runs at 30-60Hz if the docker/CPU/GPU drivers are not configured correctly.

For now you could make sure your GPU is being used (and not the on-line graphics). try nvidia-smi in the docker to see. You can also slow the dataset down using the slider in the yarpdataplayer just as a test to see if it helps for now.

I'll try to re-run the experiments when I get time - we are also trying to actively develop and remove some causes for failure so hopefully we solve this issue.

[jorou125]

Hi @arrenglover ,

Good news, I was able to run EDOPT at 200-300Hz. The communication with the GPU was not properly set. I am yet able to reproduce your results from your article (sometimes the tracking fails) but in general the tracking works well (especially when slowing down the dataset using the slider). If you could still provide me the appropriate parameters to reproduce your results I would appreciate it!