Table 9.

Results for SSL-trained models for models with various hyper-parameters on the 2016 annotated dataset together with unannotated images from harvests from 2015, 2017, and 2019. Additional results are also shown for baseline fully supervised models in the first rows.

Train Set	Unsup. Set	Bbox Thresh	Unsup Images	Unsup Weight	AP	AP@0.5	AP@0.75	PCC CW40	PCC CW43	PCC CW40+43
2016	NA	NA	NA	NA	4.97	7.39	6.05	0.70	0.54	0.56
2016	1517+2019	0.1	1	0.5	11.95	24.02	9.65	0.72	0.65	0.63
2016	1517+2019	0.1	1	4	-	-	-	-	-	-
2016	1517+2019	0.1	4	0.5	14.24	28.51	11.51	0.70	0.76	0.59
2016	1517+2019	0.1	4	4	12.00	22.43	11.32	0.74	0.66	0.65
2016	1517+2019	0.3	1	0.5	13.67	27.15	10.89	0.70	0.57	0.52
2016	1517+2019	0.3	1	4	-	-	-	-	-	-
2016	1517+2019	0.3	4	0.5	13.28	27.90	9.35	0.85	0.55	0.62
2016	1517+2019	0.3	4	4	13.53	24.15	13.31	0.84	0.66	0.70
2016	1517+2019	0.5	1	0.5	15.05	29.60	12.30	0.73	0.64	0.56
2016	1517+2019	0.5	1	4	-	-	-	-	-	-
2016	1517+2019	0.5	4	0.5	16.98	33.60	13.98	0.83	0.70	0.71
2016	1517+2019	0.5	4	4	16.62	32.75	14.16	0.79	0.65	0.58
2016	1517+2019	0.7	1	0.5	12.34	21.14	13.52	0.82	0.55	0.64
2016	1517+2019	0.7	1	4	13.92	27.88	11.91	0.85	0.61	0.58
2016	1517+2019	0.7	4	0.5	9.67	16.23	10.59	0.74	0.59	0.64
2016	1517+2019	0.7	4	4	17.66	33.91	16.24	0.90	0.61	0.64