search

dlaidig / broad

Berlin Robust Orientation Estimation Assessment Dataset (BROAD)
Creative Commons Attribution 4.0 International
28 stars 12 forks source link

NaN value in .mat file #4

Closed hoangvietdo closed 1 year ago

hoangvietdo commented 1 year ago

Thanks for sharing your incredible work!

I've witnessed some NaN values in the .mat file (e.g., dataset number 01 and 04). Is it a bug, or there are some underlying reasons for this?

dlaidig commented 1 year ago

Thanks for looking into the BROAD dataset! :)

If I'm not mistaken, the NaN values should only appear in opt_quat and opt_pos. Those values indicate short gaps in which the optical motion capture system failed to track at least one of the reflective markers. We decided to not fill those gaps using interpolation but instead exclude the corresponding samples from the error calculation. This can easily be done in Python and Matlab by using the nanmean function for the RMSE calculation:

https://github.com/dlaidig/broad/blob/6738875895e821afda12a5dd1b186092c6cff6a4/example_code/broad_utils.py#L166

Could you please confirm that the NaN values you found are indeed short gaps in the optical reference data?

hoangvietdo commented 1 year ago

Thanks for your fast response!

I indeed found the NaN value in the opt_quat. When I convert them to the Euler angle and plot them, I witness a lot of discontinuous parts due to the NaN value.

I'm fully aware of the problem with the optical system, so I can somehow guess the reason.

Thanks for your information.