Validation
What has been checked
The three analyzers were ported line-for-line to Python and run on simulated mice with known ground truth (the scripts are in Tools/validation/ in the repository).
Gait
Simulated trot, 20 cm/s, 4 Hz, duty factor 0.60, hind base of support 3 cm, 1–2 px keypoint noise:
| Measure | Truth | RodentLimb |
|---|---|---|
| Stride length | 5.00 cm | 5.00 cm |
| Stride frequency | 4.00 Hz | 4.00 Hz |
| Duty factor (100 fps) | 0.60 | 0.59 |
| Duty factor (30 fps) | 0.60 | 0.70 |
| Hind base of support | 3.0 cm | 2.98–3.00 cm |
| Step length (hind) | 25 px | 25.6 px |
| Nose lateral sway | 0.133 BL | 0.129 BL |
| Stride length @ reference speed (mixed speeds) | 50.0 px | 50.2 px |
| Phases LH→RH / LH→RF (trot) | 0.5 / 0 | 0.49–0.52 / 0.00–0.01 |
| Gait pattern (trot) | trot | 26 of 28 strides trot |
| Strides while turning at 40 °/s | rejected | 0 kept |
The stance-detection defaults (step speed = smaller of the backward and forward steps; stance ratio 0.8) were chosen from these simulations. At 30 fps, duty factor is biased by about one frame, and the app warns below 50 fps.
Tail suspension
Simulated 6-minute test with random mobile / immobile periods, breathing movement during immobility, and 0.8 px noise: immobility 163.5 s vs 165.4 s true, latency 5.7 s vs 5.6 s, 99.2 % frame-level agreement.
Cylinder test
Simulated bottom-view session with 9 scripted rears (left first, right first, both): 9 / 9 labeled correctly, also with 3 px noise and 20 % of paw detections missing.
What simulation cannot tell you
Simulations test the analysis, not the pose model on your videos. Before using the results in a study:
- Check the overlay on a few videos in the Review tab.
- Score a subset by hand, blinded, for example 10 % of animals, and compare: immobility time, contact counts, stride measures.
- Fix the parameters that best match your manual scoring, apply them to all sessions, and report them (the methods export contains them).
- For cylinder sessions, review the contact list. Corrections are stored and counted.
Report the agreement (for example a correlation or Bland–Altman plot) in your methods.