Table 3.

Performance when training and testing on different datasets.

Training data	Testing data	AbsCountDiff	CountDiff	MSE	$R^2$	Agreement (%)
Ara2013-Canon	Ara2012	5.45 (2.04)	$-$ 5.45 (2.04)	33.9	$-$ 4.79	0
Ara2012	Ara2013-Canon	5.39 (1.99)	5.39 (1.99)	33.13	$-$ 6.15	0
S12	Ara2012	1.38 (1.03)	$-$ 0.25 (1.7)	2.97	0.42	22
S12	Ara2013-Canon	1.82 (1.38)	0.46 (2.24)	5.25	$-$ 0.33	20

Training on a single dataset of synthetic rosettes performs significantly better than training on a dataset of real rosettes with a different distribution of phenotypes