search

SainsburyWellcomeCentre / aeon_experiments

Experiment workflows for Project Aeon
BSD 3-Clause "New" or "Revised" License
2 stars 0 forks source link

Aeon4 sleap model for BAA-1104048, 49 performs poorly #466

Closed jkbhagatio closed 6 months ago

jkbhagatio commented 7 months ago

@lochhh got these stats on single-animal frames after training:

BAA-1104048 Accuracy: 0.542 BAA-1104049 Accuracy: 0.872 ID accuracy: 0.706 Total tracks: 165264 Tracks identified: 165254 Tracks correctly identified: 116660

@lochhh also noted that the frames on aeon4 appear to be more blurry than aeon3 for the top cam. We may try adjusting the lens focus.

jkbhagatio commented 7 months ago

@Dario55 mentioned how opening more the aperture on the top cam on aeon4 makes it more likely the edges of the image get blurry (out of focus) relative to the center.

So in addition to the focus, we could play with decreasing the aperture opening and increasing the exposure time.

lochhh commented 7 months ago

new model is in ceph\aeon\aeon\data\processed\test-node1\4350621\2024-01-22T19-00-00\topdown-multianimal-id-133 :

single-animal frames: BAA-1104048 Accuracy: 0.926 BAA-1104049 Accuracy: 0.974 ID accuracy: 0.95 Total tracks: 213528 Tracks identified: 213521 Tracks correctly identified: 202789

jkbhagatio commented 7 months ago

Closing this as done.

lochhh commented 6 months ago

retrained model with ~200 more frames with temporal discontinuities (i.e. frames in which the diff in Euclidean distance with the previous frame > 100px)

model path: /ceph/aeon/aeon/data/processed/test-node1/4546247/2024-02-13T17-00-00/topdown-multianimal-id-133/

single animal frames: BAA-1104048 Accuracy: 0.965 BAA-1104049 Accuracy: 0.924 ID accuracy: 0.944 Total tracks: 213528 Tracks identified: 213528 Tracks correctly identified: 201653