Improved CenterNet for Accurate and Fast Fitting Object Detection

Table 1

Results on TFITS test set.

Methods		Backbone	mAP/%	FPR	FNR	Size/Mb	FPS
Single-stage detectors
SSD [19]		ResNet101	74.1	28.4	23.5	90.6	13.1
YOLOv3 [21]		DarkNet53	75.0	26.5	20.5	59.6	20.0
YOLOv4 [23]		DarkNet53	78.6	23.1	19.2	30.0	25.0
RetinaNet [51]		ResNet101	79.1	22.3	19.3	65.3	10.2
Two-stage detectors
Fast R-CNN [27]		ResNet101	77.1	22.8	18.4	98.0	8.8
Faster R-CNN [25]		ResNet101	80.2	20.0	15.3	122.0	7.0
Anchor-free detectors
ExtremeNet [32]		Hourglass104	80.9	19.6	14.7	150.6	16.8
CornerNet [31]		Hourglass104	77.9	23.2	19.8	160.0	9.2
EfficientDet [52]		EffcientNetB0	74.0	26.1	20.8	80.6	47.0
CenterNet
CenterNet-HG		Hourglass104	78.1	23.8	19.4	220.3	14.7
CenterNet-Res18		ResNet18	69.9	30.2	25.9	60.3	55.0
CenterNet-Res50		ResNet50	74.6	27.6	23.5	128.0	48.5
CenterNet-Res101		ResNet101	76.0	25.1	20.6	180.0	22.5
CenterNet-DLA34		DLA-34	76.4	24.0	20.1	77.0	27.2
M-CenterNet		MobileNetV3	75.9	25.3	20.9	18.5	43.2
HRM-CenterNet		MobileNetV3	80.3	19.8	13.6	24.6	32.5
HRM-CenterNet
HRFF	LRFEM	Backbone	mAP/%	FPR	FNR	Size/Mb	FPS
		MobileNetV3	75.9	25.3	20.9	18.5	43.2
√		MobileNetV3	79.0	22.1	16.6	24.0	33.0
√	√	MobileNetV3	80.3	19.8	13.6	24.6	32.5