Negative Y-coordinates

Hello,
I’ve been using DeepLabCut for a while, but now I’ve just trained a different network with some new videos and I’m having issues with Y-coordinates. I check the coordinates in the .csv file extracted from the analysis of the videos and some of the Y-coordinates are negative. I don’t really understand why, because (0,0) is top-left. Is that normal? Am I giving a wrong interpretation or something? Or how should I check it?

Thank you!

AFAIU …
… the ‘marker’ points are connected by the defined ‘skeleton’.

Therefore the network can predict the location of markers outside the image from markers inside the image.
All positions outside the image coordinates (x<0, x>ImageWidth,y<0, y>ImageHight) are pure predictions.
You can decide to reject them from your analysis in general. Or you can use the likelihood to define a rejection threshold.

Positions far outside the image dimensions usually have a low likelihood and should be rejected.

Thank you for your answer!

However, I’m not sure I fully understand you … Why does the network predict the location of markers outside the image from markers inside the image?

And then, do you suggest rejecting them from the analysis?

Thanks again!

To name it simple (just a hint - not the truth):

If you know where the ears are you can assume where the nose is even if it is not visible.
This prediction is based on the realtionship between ears and nose as defined by the manual classification.

@Anna_Teruel, this may also be an artefact after location refinement, which is normally corrected by default. Could you tell me if move2corner is set to True in your config?

1 Like

@jeylau Yes, it says move2corner = true

This can happen & indeed is due to the location refinement layer, that predicts body parts locations modulo the downsampled score map layer. This should just be a few pixel max.

Indeed if you want to refine the bodyparts, then “move2corner” is useful, as otherwise the points do not show up in the refinement GUI for correcting the locations.

Thank you very much for all your answers! :grinning: I do appreciate it.

However, I’m afraid I do not completley understand you… What do you mean by “location refinement layer”? Where could I read more information about this issue?

And I don’t know how to refine these labels, because likelihood values are pretty high, despite it gives me negative coordinates.

This might be helpful