You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, I'm working on analyzing the .SLP file corresponding to full-length labeled videos but continue to see that the shape of the video is being called none. In addition, when running my trained model on videos, I am noticing that redundant tracks are being created instead of continuation of existing tracks. Any thoughts on these two points? Happy to provide additional information.
In SLEAP, we do not store Video metadata such as shape and instead rely on the video being found from the path and reload that information each time. Thus, it sounds like you are doing analysis on an slp file that was moved from it's original placement (or the videos were moved), SLEAP is having trouble finding the videos, and a shape of None is returned.
However, we have been developing a complementary package sleap-io that
is intended to be a complement to the core SLEAP package that aims to provide functionality for interacting with pose tracking-related data structures and file formats with minimal dependencies. This package does not have any functionality related to labeling, training, or inference.
Eventually sleap-io will be used as a replacement to any of the data-storing classes in SLEAP. Currently it would be useful in your case due to it's feature set aimed specifically at manipulating and extracting SLEAP data (e.g. fix video paths to help your moved SLP find it's original videos - then once saved, sleap-io ensures that the video metadata such as shape is also saved to the slp).
Let us know if you need any additional help and/or example code to get started.
Re-using Tracks
In SLEAP, we have this "Max number of tracks" option that when set should re-use tracks that already exist. I've included a screenshot of the GUI option below, but this functionality is also available via the sleap-track CLI by appending --tracking.max_tracks followed by an integer for the number of tracks:
--tracking.max_tracks TRACKING.MAX_TRACKS
Maximum number of tracks to be tracked by the tracker. (default: None)
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
Hi, I'm working on analyzing the .SLP file corresponding to full-length labeled videos but continue to see that the shape of the video is being called none. In addition, when running my trained model on videos, I am noticing that redundant tracks are being created instead of continuation of existing tracks. Any thoughts on these two points? Happy to provide additional information.

Beta Was this translation helpful? Give feedback.
All reactions