Hi, I’ve been successfully labelling, training and analysing videos. But I’ve noticed that after training, when evaluating the training, the labels seem to be swapped. For example:
Here you can see the red and purple are always swapped, and the green and blue always swapped. The orange is always correct. Apart from that the inference is really good. Could anyone offer an idea why this might be happening?
Looking into this now! Thanks for raising this bug…
I just re-run the analysis (evaluation) of an old project with DLC 2.1 & DLC 2.1.6.2 and get the same pixel errors & plots. Will further look into this, but for now I do not know what the issue is. Did you train the model with DLC 2.1.6.2 or a different version?
Hi, thanks for looking into this 
This is using DLC 2.1.5.2.
For both training & evaluation?
Yes, I even cleared out all the folders except for labelling and redid the training and evaluation from scratch, under 2.1.5.2.
Ok, I also get the same pixel errors in DLC 2.15.2. The plots are also identical:
DLC 2.1.5.2:
and DLC 2.1.6.2:
One thought we had: Have you perhaps overwritten the training set that was actually used for training? I.e. dlc models contains a model that was trained with different data?