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

Table 6

Comparison of the test results of FFAD with other state-of-the-art object detectors. Results are evaluated on COCO test-dev.  ∼ indicates multiscale testing is used.

Method	Backbone	AP	AP50	AP75	APS	APM	APL
Two-stage detectors
Faster RCNN w/FPN [19]	ResNet-101	36.2	59.1	39.0	18.2	39.0	48.2
Deformable R–FCN [40]	Inc-Res-v2	37.5	58.0	40.8	19.4	40.1	52.5
Mask-RCNN [21]	ResNext-101	39.8	62.3	43.4	22.1	43.2	51.2
Soft-NMS [41]	ResNet-101	40.8	62.4	44.9	23.0	43.4	53.2
SOD-MTGAN [42]	ResNet-101	41.4	63.2	45.4	24.7	44.2	52.6
Cascade-RCNN [43]	ResNet-101	42.8	62.1	46.3	23.7	45.5	55.2
TridentDet [44]	ResNet-101	42.7	63.6	46.5	23.9	46.6	56.6
TSD [11]	ResNet-101	43.2	64.0	46.9	24.0	46.3	55.8
SNIP^∼ [1]	DCN + ResNet-101	44.4	66.2	49.2	27.3	46.4	56.9
SNIPER^∼ [29]	DCN + ResNet-101	46.1	67.6	51.5	28.0	51.2	60.5
One-stage detectors
DSSD513 [45]	ResNet-101	33.2	53.3	35.2	13.0	35.4	51.1
RefineDet512 [46]	ResNet-101	36.4	57.5	39.5	13.6	39.9	51.4
RetinaNet800 [25]	ResNet-101	39.1	59.1	42.3	21.8	42.7	50.2
PPDet [47]	ResNet-101	40.7	60.2	44.5	24.5	44.4	49.7
AutoFPN [48]	ResNet-101	42.5	-	-	-	-	-
FreeAnchor [49]	ResNet-101	43.0	62.2	46.4	24.7	46.0	54.0
M2Det ^∼ [7]	ResNet-101	43.9	64.4	48.0	29.6	49.6	54.3
FoveaBox [50]	ResNext-101	42.1	61.9	45.2	24.9	46.8	55.6
FCOS [26]	ResNext-101	44.7	64.1	48.4	27.6	47.5	55.6
CornerNet [51]	Hourglass-104	40.6	56.4	43.2	19.1	42.8	54.3
ExtremeNet [52]	Hourglass-104	40.1	55.3	43.2	20.3	43.2	53.1
CenterNet [53]	Hourglass-104	44.9	62.4	48.1	25.6	47.4	57.4
CenterNet ^∼ [53]	Hourglass-104	47.0	64.5	50.7	28.9	49.9	58.9
RepPoints [54]	DCN + ResNet-101	45.0	66.1	49.0	26.6	48.6	57.5
Ours
FFAD	ResNet-101	44.1	62.2	47.9	27.4	47.6	56.7
FFAD	DCN + ResNet-101	46.5	64.9	51.2	29.3	51.3	60.8
FFAD	DCN + ResNext-101	47.4	66.9	52.0	31.1	51.5	61.9
FFAD ^∼	DCN + ResNext-101	49.5	68.9	53.9	35.8	53.6	63.3