A Novel Pyramid Network with Feature Fusion and Disentanglement for Object Detection

Table 4

Comparison of detection AP results of different architectures.

Method	Testing time (ms)	AP	AP50	AP75	APS	APM	APL
RetinaNet	56	35.7	55.0	38.5	18.9	38.9	46.3
RetinaNet + FFAD	58	37.5	57.1	40.4	22.7	41.2	49.3
FCOS	45	37.1	55.9	39.8	21.3	41.0	47.8
FCOS + FFAD	48	39.7	58.2	43.5	24.7	43.5	52.2
ATSS	44	39.3	57.5	42.8	24.3	43.3	51.3
ATSS + FFAD	49	41.4	59.1	45.0	25.1	45.5	53.8

All models were trained using ResNet-50 backbone and the same training strategies. Results are evaluated on COCO minival set.