Cannot extract feature data or view plots (best parameters for low resolution videos?)

[willgryan]

Hi there!

I'm trying to use Tierpsy Tracker to track multiple worms in a 640x480 recording. While I've been able to get good masks of the worms and bounding boxes surrounding them, I'm unable to extract any feature information of them, and in Viewer I'm unable to access any plots of the data.

Below is an image of what I'm describing:

The main parameters I've adjusted are not using the default pytorch NN (per reading other issues with similar problems), along with background subtraction, threshold, and dilation. I suspect this has to do with the low resolution nature of the recordings, but I'm not sure if the inability to extract feature information or view plots is because of parameters or user error.

Thank you!!

Sample video and parameters here.

[luigiferiani]

Hi @willgryan,

The feature information and plots should be available if you select a *_featuresN.hdf5 file in the Select Skeletons File. When you click Select Skeletons File you'll need to select Files of type: All files (*), then you'll be able to select the _featuresN.hdf5 files.

In alternative, you can edit the filename in the Select Skeletons File to end in _featuresN.hdf5, then press the Enter key to apply the change.

I understand this is a little cumbersome, but we can make it less so with the next update.

The easiest way at the moment is to select the masked video first, that will automatically select the *_featuresN.hdf5 file

Let me know if this fixes your issue,

Cheers, Luigi

[luigiferiani]

That was an even quicker fix than I thought it would be. If you update Tierpsy, when clicking on the Select Skeletons File, featuresN and features file will not be greyed out any more :)

[willgryan]

@luigiferiani

Thank you! This has resolved my issue :)

I read that improved support for low-res videos was on the agenda for the Tierpsy Tracker hackathon that took place previously. Did anything come of this? I.e, perhaps optimal parameters for dealing with these types of recordings. I noticed that worm trajectories 'get lost' when they come into close proximity then separate again, and are treated as new worms. My initial thought would be reducing the dilation level so masks don't interfere as much, but perhaps you would know better.

Thank you!

Best,

William Ryan

[luigiferiani]

Hi - Tierpsy at the moment can't really handle worms touching each other. We're working on that too! But it's not really the dilation masks touching that starts the issue - see screenshot here:

There are definitely cases in which two worms are not touching each other yet, but Tierpsy thinks they do, and skeletonisation fails - this is due to non-perfect binarisation of the worm's region of interest. A tell-tale sign that this is happening is that the worm contours, as found by Tierpsy, don't follow closely the actual worm (they're too "fat"). Setting the parameter "worm_bw_thresh_factor" to 1.1 usually helps in these cases. Sorry, that did not seem to work on your video.

I'll see if I can tune your parameters a little though, that's a tricky video :)

Lastly, opening the featuresN.hdf5 result file should plot 'smoothed' skeletons - the result will be better than plotting the 'raw' skeletons you get from the skeletons.hdf5 file