Table 3.

Comparison between gold standard (GS; expert classifications), citizen science (CS; Penguin Watch) and computer vision (CV; Pengbot) counts (n = 1183).

	GS vs. CV (combined)	GS vs. CV (adults)	CS vs. CV (combined)	CS vs. CV (adults)	GS vs. CS (combined)	GS vs. CS (adults)
DAMOa	n = 300
Average difference	1.95	2.01	2.21	2.25	1.36	1.18
σ	1.84	2.15	2.16	2.38	1.66	1.56
Proportion 0 or 1	0.49	0.49	0.47	0.47	0.70	0.75
Overestimate	121	143	116	145	102	101
Underestimate	115	90	136	102	113	100
HALFc	n = 283
Average difference	1.50	1.21	1.46	1.45	1.20	0.88
σ	1.36	1.08	1.70	1.56	1.53	1.34
Proportion 0 or 1	0.61	0.69	0.65	0.66	0.72	0.85
Overestimate	60	107	70	109	58	64
Underestimate	161	101	134	99	116	91
LOCKb	n = 300
Average difference	1.83	3.71	2.13	3.50	1.23	1.17
σ	1.70	4.27	2.50	4.10	2.03	1.98
Proportion 0 or 1	0.55	0.45	0.52	0.46	0.72	0.77
Overestimate	68	157	80	159	61	82
Underestimate	168	96	151	81	92	80
PETEc	n = 300
Average difference	12.13	7.17	11.31	7.09	3.70	2.36
σ	5.07	3.39	7.19	4.17	4.12	2.69
Proportion 0 or 1	0	0.06	0.02	0.08	0.36	0.46
Overestimate	0	81	1	84	85	79
Underestimate	300	212	297	207	170	162

Average differences (in raw count) are provided, alongside the standard deviation of these differences, the proportion of counts that were equal or differed by only one penguin, and the number of over- and underestimates. For GS and CS, combined counts (i.e. adults and chicks) and adult-only counts are included. The filtering threshold levels used for the CS counts are num_markings > 3 for adults and num_markings > 1 for chicks. CV cannot distinguish between adults and chicks, so a single value is provided. Overestimates and underestimates relate to the second variable, i.e. for DAMOa ‘GS vs. CV (combined)’, computer vision overestimated the penguin count in 121 cases, and underestimated it in 115 cases, as compared to the gold standard. GS vs CS counts are included for completeness – please see Jones et al. (2018) for a full discussion.