DeepSeedling: deep convolutional network and Kalman filter for plant seedling detection and counting in the field

Table 1 Overall performance of the model generalizability

Training set	Testing set	mAP \(\text {IOU}_{\text {all}}\)	mAR100 \(\text {IOU}_{\text {all}}\)	F1 \(\text {IOU}_{\text {all}}\)	mAP \(\text {IOU}_{\text {0.5}}\)	mAR100 \(\text {IOU}_{\text {0.5}}\)	F1 \(\text {IOU}_{\text {0.5}}\)
T15U15 training	UGA2018 testing	0.763	0.808	0.785	0.981	1.000	0.991
UGA2018 training	UGA2018 testing	0.778	0.818	0.798	0.983	0.992	0.987
T15U18 training	UGA2015 testing	0.179	0.335	0.233	0.352	0.683	0.464
UGA2015 training	UGA2015 testing	0.599	0.695	0.643	0.923	0.997	0.959
U15U18 training	TAMU2015 testing	0.377	0.535	0.442	0.588	0.857	0.698
TAMU2015 training	TAMU2015 testing	0.791	0.827	0.809	0.989	1.000	0.995

For the testing set in each of the UGA2018, UGA2015, and TAMU2015 datasets, two Faster RCNN models were trained using the training set from the same dataset and from the combination of the other two datasets, respectively. Performance comparison of the two models was used to evaluate the model generalizability

ISSN: 1746-4811