HOTA

Simplified, not a 28 page paper on a metric @-@

HOTA = Higher Order Tracking Accuracy.

I will take a top down approach to this metric, which is more intuitive.

Above shows how HOTA is the geometric mean of how well you detect (DetA = Detection Accuracy) and how well you link identities over time (AssA = Association Accuracy). One number which is equally weighted between the two.

Now here is how those 2 are calculated:

Our detection accuracy is all our true positives (meaning we predict a point/box on an object, where the object is truthfully present), divided by the sumation of our true positives with our false negatives & false positives. False negative being, if there is a true object present, but we make no predictino on it, and a false positive being if we make a prediction where there is no object present.

Our association accuracy is the average association accuracy.

is a Haccard index over trajectories. For a matched detection c with predicted ID p and ground-truth ID g:

• TPA(c) : all other matched deetections in the video that also pair p with g
• FNA(c) : detections with ID g that the tracker gave some other ID (or missed)
• FPA(c) : detectinos the tracker labbeled p that actually belong to some other object

Every true positive is looking through the whole track, to see how many it got right. Which is the temporal question, and why the gemoetric mean is meaningful in this contect.

HOTA is a more modern mertic used for tracking. The previous default metric, MOTA (Multi-Object Tracking Accuracy), had an issue of over-weighing our detection; while IDF1 (IDentification F1) over-weights association. HOTA splits them cleanly, so a score drop tells you which half of your tracker is failing, and the sub-metrics are individually reportable.