Table 2.

Comparison of algorithms on the apple dataset. The best-performing experimental data in the table are represented by bold numbers.

Method	Backbone + Neck	AP	AP0.5	AP0.75	APs	APm	APl	AR	ARs	ARm	ARl
Two-stage anchor-based
Faster R-CNN	ResNet50 + FPN	59.6	85.9	65.8	43.6	67.0	84.3	65.5	51.5	72.4	88.3
Mask R-CNN	ResNet50 + FPN	60.1	86.3	66.5	44.9	67.4	84.9	66.4	53.2	73.0	88.2
MS R-CNN	ResNet50 + FPN	60.2	86.3	67.3	45.3	67.1	84.9	66.4	53.6	72.6	87.9
One-stage anchor-based
RetinaNet	ResNet50 + FPN	57.6	84.9	62.6	42.2	65.1	82.8	65.1	50.8	72.6	87.5
YOLOv3	DarkNet53	59.4	84.6	65.2	40.8	65.9	87.5	65.9	51.6	71.9	91.7
ATSS	ResNet50 + FPN	62.1	87.9	64.7	46.1	67.2	88.8	69.3	56.4	75.0	92.0
One-stage anchor-free
FoveaBox	ResNet50 + FPN	58.6	86.2	63.8	43.8	63.6	83.5	66.6	54.7	72.0	87.2
FCOS	ResNet50 + FPN	58.4	86.8	63.4	42.6	64.3	83.7	65.6	51.7	72.9	87.9
Ours
FCOS-LSC	ResNet50 + FPN + LSC	63.0	87.2	68.1	47.1	69.5	89.9	71.3	58.5	77.3	92.7