Table 3.

Training protocol values.

	Hyper-Parameters
Method	Input Size	$\mathbf{Lr}$	Epochs	Batch Size	${\mathbf{Lr}}_{\mathbf{decay}}$	Post-Processing	Confidence Threshold
EfficientDet-d0 ${}^1$	$512\times 512$	$0.08$	300	48	200, 250	Soft-NMS	$0.4$
EfficientDet-d5 ${}^1$	$1280\times 1280$	$0.08$	300	12	200, 250	Soft-NMS	$0.4$
Faster R-CNN ${}^2$	$800\times 800$	$0.0001$	300	8	243	NMS	$0.5$
Mask R-CNN ${}^2$	$800\times 800$	$0.0001$	300	8	243	NMS	$0.5$
RetinaNet ${}^2$	$800\times 800$	$0.0001$	300	8	243	NMS	$0.5$
YOLO-v5x ${}^3$	$640\times 640$	$0.0032$	100	12	–	NMS	$0.001$
YOLO-v5s ${}^3$	$640\times 640$	$0.0032$	100	12	–	NMS	$0.001$

¹https://github.com/google/automl/tree/master/efficientdet (accessed on 2 July 2021). ² https://github.com/facebookresearch/detectron2 (accessed on 2 July 2021). ³ https://github.com/ultralytics/yolov5 (accessed on 2 July 2021). Lr—learning rate. Lr_decay—epochs in which the learning rate is decayed.