How to read
% deviation = |predicted − actual| / actual × 100; Δ count = predicted − actual. Mean deviation weights images equally, excluding missing predictions and zero actual counts. Confidence is P(exact count); chart accuracy requires an exact match.
Header arrows sort models; model-row arrows sort images. Provider grouping keeps sorts within providers. Missing values stay last. Cost/time means exclude missing values and skipped calls; totals include known failed-call charges. * marks a partial total. Times are summed request latencies, not wall time.
Latest record wins per image/configuration. Totals cover displayed records. Charts omit invalid confidence and ambiguous ground truth; coverage includes all images with known ground truth. Equal-confidence answers share one cutoff.