Not-visible nodes not hidden in csv

[beeldash]

Hello all,

I am new to SLEAP and am having some trouble exporting my predictions as a CSV. I am using the GUI for all steps.
Each frame contains only one animal (i.e., one instance), and I am tracking four nodes using a multi-animal top-down model.
I ran inference on the entire video, then went to Labels -> Accept All Predictions and manually corrected the node positions when needed. However, when I export the analysis via File -> Export Analysis CSV, I encounter two issues:

1. Non-visible nodes are not marked as NaN - I get the coordinates of the hidden node instead (e.g. screenshot of 1st frame)
2. Most, but not all, frames contain two rows that are nearly identical, except for the score value.

Am I missing something here?
Thank you!

[tk1475]

Both of these come from how the CSV exporter works. You didn't do anything wrong.

Duplicate rows. Accept All Predictions doesn't replace the predictions. It adds a user instance next to each one and links it back through from_predicted (AddUserInstancesFromAllPredictions). The GUI hides the linked prediction, but the CSV export (CSVAdaptor.write → sleap-io _to_instances_df) writes one row for every user instance and one for every predicted instance (dataframe.py). In your screenshot, the row with empty scores is your corrected user instance, and the row with scores is the original prediction. Frames with a single row have only one of the two.

Hidden nodes. The same function writes point["xy"] for every node and never checks the visible flag (L1128-L1134). Marking a node as not visible keeps its coordinates, so they end up in the CSV.

What to do:

1. Use File → Export Analysis HDF5 instead. That path calls inst.numpy(), which turns non-visible points into NaN by default (numpy.py#L475, instance.py). In single-instance projects it also drops predictions on frames that have a user instance (numpy.py#L84-L92). You can read the result with h5py or pandas.
2. If you need the CSV, run Labels → Delete Predictions on User-Labeled Frames... first and pick "All predictions on user-labeled frames". The default "Unlinked only" option keeps linked predictions like yours (commands.py). You'll still need to blank the hidden nodes yourself afterwards.

The CSV ignoring visible looks like a bug, so it may be worth opening an issue on sleap-io.