Train sleap id model on patch-cam views - Githubissues

SainsburyWellcomeCentre / aeon_experiments

Experiment workflows for Project Aeon

BSD 3-Clause "New" or "Revised" License

2 stars 0 forks source link

Train sleap id model on patch-cam views #490

Open jkbhagatio opened 5 months ago

lochhh commented 5 months ago

note: #486

lochhh commented 5 months ago

curate 100? single-animal frames per patch, per animal using RFID detection timestamps as sleap.gui.suggestions.SuggestionFrames
use existing non-ID patch cam model (e.g. Y:\aeon\code\scratchpad\sleap\multi_point_tracking\single_animal_CameraPatch\models\231201_173729.single_instance.n=400) to predict body parts without ID
assign "Tracks" to these predictions and convert to "user-labelled" frames

We should have 600 labelled frames (2 animals, 3 patches) for training patch ID model.

Using the patch ID model, infer on multi-animal frames for prediction-assisted labelling. Multi-animal frames can be selected:

as in step 1. above,
by finding large blobs from CameraTop.Position, (i.e. both animals are at the same patch), and/or
using CameraTop.Pose to find the times when both animals are at the same patch