Negative values for pose_Tz

[NicolasStefaniak]

Hello,

with one of my colleague, we are using open-face. When we have analyzed data for pose_Tz, it appeared that we have some negative values, while it is not possible that the distance between the participant and the screen is negative.

OpenFace was running on a Windows 10 (education) computer, version 22h2
We have identified this problem for 5 values among about 6 millions.

Note : for one of the participants, she moved her arm before her eyes and her pullover is black. Nevertheless, there is only one frame with a negative value. We have attached the file in order to help to understand (see for instance Frame 8888).

Best regards,
Nicolas
6xF92Twy_Peur.csv

[brmarkus]

Would you see such outliers happening with a sequence of frames (result of a certain history of samples), or could you reproduce it with a specific, single frame?

Do you get such negative values if you only feed-in the frame 8888?

Not sure about whether the maintainers are still actively watching this repository...
Maybe someone could have a closer look if you could provide a sample frame?
Was there fast movement - or the post close to a "flip point" (euler coordinates)?

In general - "big data" might anyway get processed and being "robust enough" against outliers... applying a filter...

[NicolasStefaniak]

Dear brmarkus,
thank you for your response. I am going to test the frame but I am not sure I will be able to identify the frame in the video accurately.
There was a fast movement. It may be the origin of the problem.
I cannot share the video due to the GDPR.
Filter is also the solution that seems to us the better one.
Best,
Nicolas