Fig. 4

Boxplots of performance (F1 score) on the testing set for models initialized using different pretrained models. a Are results for the UGA2018 dataset using models initialized by weights pretrained on the COCO and T15U15 datasets, respectively, b are results for the UGA2015 dataset using models initialized by weights pretrained on the COCO and T15U18 datasets, respectively, and c are results for the TAMU2015 dataset using models initialized by weights pretrained on the COCO and U15U18 datasets, respectively. Base indicates model initialization using weights pretrained on the COCO dataset, whereas DA indicates model initialization using weights pretrained on a domain dataset. For each of the TAMU2015, UGA2015, and UGA2018 datasets, a subset of 100 images were randomly selected from the training set to train a Faster RCNN model. A total of 10 models were obtained through 10 training repetitions for statistical comparisons between the models. Asterisks indicate statistical differences in model performance at the significance levels of 0.05 (*), 0.01 (**), and less than 0.001 (***)