Keypoint Identity Mismatch

[stofe95]

It seems the labelling GUI is sporadically switching the identity of keypoints ex: In the example, the 'Keypoint selection' indicates the orange point is "tailbase", whereas the 'Color scheme reference' indicates that orange is "back left". I had labelled 9 other folders where the 'Keypoint selection' matched the 'Color scheme reference', but then for this folder it got mixed up. When I exited (without saving) and reopened the folder, the color scheme and keypoint selection were in agreement again.

[stofe95]

In addition, I managed to label the frames and train, but then went to extract outliers and now I also have this behaviour: It seems that some frames have two sets of labels (the small dots are machine labels, the large dots are in the collected data layer). I thought this might happen if outlier extraction pulls out a frame that was also labelled manually, but I don't think this frame had been labelled manually. The collected data labels also show the wrong colours as per my original issue.

[jeylau]

Hi @stofe95, could you tell me what version of napari-deeplabcut is installed on your machine?

[stofe95]

Hi I had this issue on two machines, using 0.1.1 and 0.1.2

[stofe95]

As for my second point, reading the documentation seems to suggest that in cases where I see two sets of points, I should hide/trash the machine labels. So that is working as intended (though the identities of the collected data labels is still mismatched). Is that right?

[jeylau]

Hiding the machine labels is the correct thing to do; I'll look at the mismatch between the label color and the color scheme reference though, this looks like a bug.

[jeylau]

@stofe95, I could not reproduce the color bug. Does this happen while labeling or upon loading some data?

[stofe95]

I think that it happens both when you are freshly labelling new data, if you are labelling keypoints A, B, C, if on the first frame of labelling keypoint A isn't in the frame and so you select keypoint B and label it, then the colours get mixed up. This I managed to get around by exiting without saving, then reloading the data and even if the keypoint was not visible, placing the label for all keypoints on the first frame, but the ones that weren't visible. It then also happens after extracting outlier frames when you load the data and is immediately there and I don't know how to fix that.

[jeylau]

@stofe95, what you described was actually fixed recently. Try to update napari-deeplabcut to use the very latest version (still unpublished to pypi) with pip install git+https://github.com/DeepLabCut/napari-deeplabcut.git: this will install the latest fixes 😊