This protocol takes you from a YORU model and a few videos to an analysed table of per-animal tracks. The example is a pair of flies in a courtship chamber, detected with the classes solo and copulation — but the steps are the same for any animal.
Follow the Install guide, and check that it starts:
cd YORU-dev\YORU-Tracker
uv run yoru-tracker --version
YORU Tracker has no detector of its own; it uses the model you trained in YORU (.pt, .pth or .onnx). If you do not have one yet, follow YORU’s Step-by-step Protocols, which train a model on the Fruit Fly Copulation Dataset.
The model should detect each animal: every animal in every frame, as its own box. How the boxes are classed (by behaviour, by sex, …) matters less, but decides one setting in Step 3.
Put the videos of one experiment into one folder, for example:
D:\experiment\videos\
├── pair01.mp4
├── pair02.mp4
└── pair03.mp4
| Number | Why | Setting |
|---|---|---|
| How many animals are in view, if it never changes (here: 2) | The single most effective setting: IDs can then never exceed it. | Number of animals (population) |
| How far an animal moves in one frame at most, in pixels | Detections further than this from a track’s prediction cannot be matched to it. | Max distance (max_distance) |
| Whether classes are behaviours or animals | Behaviours (solo, copulation): an animal changes class but not identity. Animals (male, female): a class never changes. |
Match within class only (class_aware) |
To measure distances in pixels, open a frame in any image viewer and read the coordinates of a fly at two moments, or look at the vx/vy columns of a first tracking run (Step 2).
YORU’s own video analysis files are named after the video (pair01.mp4 → pair01.csv). If you already have them, or real-time recordings, see Step 5: they can be tracked without running the detector again.