Table 4.

Experimental results on both GrapeCS-ML and our internal dataset. The detector has been trained in three different ways: using the entire set 1 as train, with dataset augmentation; using only 10% of set 1 as a train, with dataset augmentation; using only 10% of set 1 as a train, without dataset augmentation.

	mAP
Dataset Name	Train Complete, with Augmentation	Train 10%, with Augmentation	Train 10%, without Augmentation
Validation (Set 2)	93.97%	90.95%	85.24%
Test (Set 3 + Set 4 + Set 5)	92.78%	90.98%	87.65%
Set 3	98.77%	98.69%	97.30%
Set 4	89.18%	86.70%	83.40%
Set 5	85.64%	80.07%	68.44%
Internal Dataset	89.90%	86.41%	70.75%