Retrieving hand-labeled points

[ndicola]

I am attempting to export the hand-labeled points for use in my data cleaning step, but these points are not included when I export to an hdf5 file. I have not see any way to do this online, so I have two methods which I'm considering which I could use some guidance on if possible.

1. Name all hand-labeled tracks a certain track name (eg rat1-hand, rat2-hand), that way when I export I know to pull those instances with those track names for all hand labeled frames. But it would mean going through and re-labeling ~4000 instances, which is not ideal. Right now, these instances don't have a track name and the instance list for a frame would look like this:
   
   Where the instances without any track names are the hand-labeled ones. Is the fact that there is no track associated with the hand-labeled instance the reason it doesn't get exported?
2. Try and extract the hand-labeled frames straight from the 'slp' file. From what I can tell, the '/point' dataset is all of the hand-labeled points (based on the number of positive 'complete' values), while '/pred_points' is obviously all the predicted points. However, references in the '/instances' dataset only seem to have ids for the predicted points, based on the fact that the maximum id values and the number of predicted points are the same. But there obviously must be a back reference to the hand-labeled points since the program would have to know which hand-labeled nodes go to which instances/frames/videos, or else it wouldn't be able to display that instance in the GUI. Am I missing something obvious in how these points are referenced by SLEAP? It doesn't seem like the instance class prohibits having an instance be hand-labeled, particularly with the "from_predicted" flag, but all of of my instances in the "/instance" dataset of the slp file reference predicted instances and no hand-labeled ones.

Thank you in advance for your help.

[ndicola]

Okay I think I figured it out. The flag for whether an instance is hand-labeled or predicted is actually in the 'instance_type' field of the '/instance' dataset saved in the raw .slp file, where '0' is hand-labeled and '1' is predicted. I thought the value would be in the 'from_predicted' field I only ever get '-1' for the attribute. However, the 'instance_type' flag doesn't seem to be kept when exporting a video's data to an hdf5 file since those instances are usually not assigned a track. This means that if I can make sure to keep the videos, frames, and instances sorted in the .slp file, I can just use that to analyze my data. This is a bit of a pain, but I can see why trackless instances wouldn't be exported. Can I please get some verification that this is, in fact, how the .slp and exported data works in case I'm missing something?

[talmo]

Hi @ndicola,

Sorry, looks like this fell through the cracks here.

I would not recommend parsing the SLP file directly -- we normally export it to Analysis HDF5 or CSV (File menu or via the sleap-convert command) as the formats are a bit more self-descriptive.

If you need to work the SLP files directly, we recommend loading and working with them using sleap-io which is a standalone library for working with SLP and other formats more natively.

If you really want to use your own parsing routines, you can check out how sleap-io implements the reading here.

Let us know if you have any questions!

Cheers,

Talmo