Acute multianimal tattooed tail sessions

[jkbhagatio]

Following up from the progress in #141, we are starting to run three-animal sessions in B2-210 (AEON3) with tattooed tails for which we will test the inference of a SLEAP trained network. We will run these sessions under the 'presocial0.1' experiment.

The plan is to run 30-60 minute sessions twice a week (Tuesdays and Fridays) with all three animals for the next 4 weeks, and assess the stability of the SLEAP trained network's inference over time.

The animals used for these sessions are:

• BAA-1103351 (dual band tattoo)
• BAA-1103352 (no tattoo)
• BAA-1103353 (single band tattoo)

The first sessions occurred on 2023-06-16.

BAA-1103353 had to be culled due to tail complications.

[lochhh]

The three-animal sessions are restarted with a new single-band-tattoo mouse and run under the 'multianimal-test' experiment in B2-210 (AEON3).

The animals used for these sessions are:

• BAA-1103351 (dual band tattoo)
• BAA-1103352 (no tattoo)
• CAA-1120139 (single band tattoo)

The first sessions occurred on 2023-07-04.

Again, CAA-1120139 had to be culled due to tail complications.

[lochhh]

The three-animal sessions are restarted the third time with a new single-band-tattoo mouse and run under the 'multianimal-test' experiment in B2-210 (AEON3).

The animals used for these sessions are:

• BAA-1103351 (dual band tattoo)
• BAA-1103352 (no tattoo)
• BAA-1103369 (single band tattoo)

The sessions occurred on

• 2023-07-28
• 2023-08-01
• 2023-08-03
• 2023-08-08
• 2023-08-11

[lochhh]

Training dataset consists of:

• 1120 single-animal frames from 2023-07-04 for BAA-1103351 (dual band tattoo)
• 1008 single-animal frames from 2023-07-04 for BAA-1103352 (no tattoo)
• 1126 single-animal frames from 2023-07-28 for BAA-1103369 (single band tattoo)
• 287 multi-animal frames from 2023-07-28

Resulting in models:

• AEON3B_pattern_tattoo3_1_topdown_top.centroid
• AEON3B_pattern_tattoo3_1_topdown_top.centered_instance_multiclass

Performance on 100 consecutive and randomly selected ground truth multi-animal frames from different sessions

Stats/Dataset	2023-07-28	2023-08-01	2023-08-03	2023-08-08	2023-08-11
Tracks identified	261	298	275	260	242
Tracks correctly identified	248	297	274	252	223
Total tracks	300	300	300	300	300
ID accuracy	0.9502	0.9966	0.9964	0.9692	0.9215
Frames with < 3 tracks	0.39	0.02	0.24	0.40	0.57
Frames with 2 missing tracks	0.00	0.00	0.01	0.00	0.01
Frames with 1 missing track	0.39	0.02	0.23	0.40	0.56
Track occupancy	[0.87 1. 0.74]	[1. 0.98 1. ]	[0.79 1. 0.96]	[1. 0.82 0.78]	[0.96 0.65 0.81]

Performance on 100 non-consecutive and randomly selected ground truth multi-animal frames from different sessions

Stats/Dataset	2023-07-28	2023-08-01	2023-08-03	2023-08-08	2023-08-11
Tracks identified	233	243	259	256	254
Tracks correctly identified	217	231	248	224	227
Total tracks	300	300	300	300	300
ID accuracy	0.9313	0.9506	0.9575	0.8750	0.8937
Frames with < 3 tracks	0.59	0.49	0.38	0.42	0.43
Frames with 2 missing tracks	0.08	0.08	0.03	0.02	0.03
Frames with 1 missing track	0.51	0.41	0.35	0.40	0.40
Track occupancy	[0.82 0.84 0.67]	[0.84 0.75 0.84]	[0.94 0.83 0.82]	[0.92 0.83 0.81]	[0.91 0.83 0.8 ]

[lochhh]

flowchart TD
    raw_data(AEON Data) --> sa_frames[/single_animal_frames/]
    raw_data --> ma_frames[/multi_animal_frames/]
    sa_frames --> |all frames| proc_auto_label[Auto labelling]
    ma_frames --> |curate| proc_manual_label[Manual labelling]
    proc_auto_label --> |curate| training_dataset[/training_dataset.slp/]
    proc_manual_label --> training_dataset
    proc_auto_label --> |all frames| sa_gt[/single_animal_groundtruth.slp/]
    sa_gt --> proc_predict
    training_dataset --> proc_train_model[Train model]
    proc_train_model --> exported_model[/model.pb/]
    exported_model --> proc_predict[Predict]
    proc_predict --> ma_pr[/multi_animal_predictions.slp/]
    proc_predict --> sa_pr[/single_animal_predictions.slp/]
    ma_frames --> |all frames| proc_predict
    sa_pr --> proc_eval[Evaluate]
    ma_pr --> proc_eval
    proc_eval --> retrain{Is model good enough?}
    retrain --> |yes| exit_loop(End training)
    retrain --> |no| proc_correct[Extract and correct random n low-scoring frames]
    proc_correct --> corrected_gt[/low_score_groundtruth.slp/]
    corrected_gt --> |merge| training_dataset

Loading