Defining confidence threshold for labeling

deeplabcut
#1

My lab is tracking many labels. As a result, several labels are always incorrect. It is not a big deal (we’ll retrain). However, we would like to set a higher confidence threshold. If uncertain, we’d prefer not to have a label at all. Does anyone know where to set this threshold value?

#2

In the config file there is a p-cutoff option, under # Plotting configuration. You can set that higher so that the labels with likelihood lower than that don’t get plotted in create_labeled_video. It is by default 0.1.

1 Like
#3

That’s what I thought. I have tried it with a higher value (0.3) but with similar results. I will look through the accompanying csv file and pick a better value. Thank you for your reply.

#4

I would go with 0.9 …