← Back to slide 27The coverage table, drawn rather than counted
← All interactives
Coverage Wheels
Pick a training set and a positional cell. Radius = magnitude band (3–300 px, log), angle = direction, deep blue = at or below the 5% deficiency bar, orange = surplus. One wheel per arm, drawn from the actual 1,728-cell counts.
shade = share of that position's 192 cells that are deficient
hover or tap a wedge to read it
5% of uniform · ¼× · uniform · 4× · 16×
unobservable: out-of-frame motion
Benchmark occupancy wheels: what the evaluation sets actually containsame binning
Occupancy of each benchmark's own annotated motion, mean over the 9 positions. The doctor never looks at these; they are shown here so you can see which holes in a training set matter for which task.
Colour is each cell's count relative to the whole-table uniform share (total ÷ 1,728), on a log₂ scale from 5% of uniform (deep blue) to 16× (deep orange). A base is mostly deep blue: hollow. The prescription asks for exactly those wedges and the fulfilled arm fills them; own-source real data densifies where mass already exists; FlyingThings additions land in FlyingThings' habits. Header numbers are the trained matcher's SEM / GEO accuracy for that arm (PCK@3%, seed means).