Palm size and shape decide which canonical finger poses the system will even see

What it is

A counting recognizer holds a set of canonical shapes: palm width, finger-length ratios, and inter-finger angles taken from the majority hand in the training data. Real hands fan out in size, ratios, syndactyly, and thickness. A child's small palm and relatively short fingers make the gaps for three and four narrower in pixels. A wide, short-fingered hand showing four may look fused on the ulnar side. The count the system reports is “which canonical shape this looks like,” not “how many fingers this hand actually extended.” When template and morphology mismatch, errors pile up on one kind of hand while mean accuracy still looks fine.

Why it happens

Visual counting depends on gaps being separable in the image. Gap width in pixels is roughly a function of palm width and distance: same count, smaller hand, farther away, and the gap smears first. If training is mostly young adults with medium palms, the decision boundary sits on that geometry. Partial syndactyly means “four” never has a full extra gap. Arthritic deviation makes extended fingers cross in projection. Depth maps drop points on dark skin and dark gloves, and gap contrast falls another step. None of this is the user “showing it wrong.” The canonical shape assumed an average hand that does not exist. Body-scale skeleton variation is a different layer; this is geometry on one effector.

Studying it

Stratify by measurable geometry: palm width, middle-finger length, syndactyly, age band. Report 0–5 confusion per stratum. Train templates on one group versus mixed groups and watch asymmetric error on small-hand and large-hand strata. Distance as a second factor: the same hand in near field and at two metres. Material conditions: dark gloves, rings. Do not report only overall accuracy. File failures by morphology, not under “non-standard users.”

Where it stops holding

Near-field headsets with millimetre tracking shrink the effect of size; syndactyly and deviation remain. Contour-only, no-depth pipelines are more sensitive to palm thickness. A prosthetic hand whose gaps are decorative and fixed may always report the same count. Thermal or capacitive gloves are not visual gaps, so this mechanism does not apply; it becomes a matching problem between electrode layout and hand shape. When children and adults share one threshold, do not let the adult stratum's success vouch for the child stratum.

Applying it

• Collection and acceptance must cover children, small palms, wide palms, and syndactyly, published in bins. Passing on colleagues' hands is not a gate.
• For morphologies that still fail up close, offer a calibration (the user extends 1–5 so the system sees their gaps) rather than one more global threshold.
• Failure copy should not offer only “please spread your fingers.” Check first whether the template covers this hand, then decide calibration or another channel.

Related

• Same group: C4.09.1 Extended-finger count can express a choice or a quantity · C4.09.3 Holding a finger combination has extra postural cost · C4.09.4 Recognition and bodily feel must be logged by hand morphology
• Adjacent: C4.13 Full-body pose and skeleton tracking · C4.04 Static poses versus dynamic gestures
• Search: hand morphology · canonical hand pose · finger segmentation