Wireless Communications and Mobile Computing

Research Article

EAWNet: An Edge Attention-Wise Objector for Real-Time Visual Internet of Things

Table 4

Item FPS and AP of different object detectors.


Method	Backbone	Size	FPS	AP (%)	AP50 (%)	AP75 (%)	APs (%)	APm (%)	APl (%)

YOLOv4: optimal speed and accuracy of object detection [19]
YOLOv4	CSPDarknet-53	416	30	35.6	57.8	38.2	17.3	39.2	52.1
YOLOv4	CSPDarknet-53	512	22	37.9	60.0	41.9	19.8	41.5	49.8
YOLOv4	CSPDarknet-53	608	16	38.2	60.9	42.5	21.7	42.1	47.4

Learning rich features at high speed for single-shot object detection [15]
LRF	VGG-16	300	39.0	26.8	46.7	29.4	8.3	30.3	42.6
LRF	ResNet-101	300	36.2	29.2	50.0	32.2	8.6	33.2	45.9
LRF	VGG-16	512	27.9	31.9	51.7	33.8	14.7	35.4	44.3
LRF	ResNet-101	512	19.7	32.6	53.2	35.1	15.2	38.0	45.4

Receptive field block net for accurate and fast object detection [20]
RFBNet	VGG-16	300	32.0	25.3	44.5	27.1	6.9	27.2	41.3
RFBNet	VGG-16	512	22.5	29.2	50.1	31.2	11.7	32.4	42.5
RFBNet-E	VGG-16	512	17.3	29.6	50.7	31.5	12.9	31.8	42.7

YOLOv3: an incremental improvement [18]
YOLOv3	Darknet-53	320	27	23.3	46.8	25.1	7.2	25.8	38.4
YOLOv3	Darknet-53	416	23	26.4	50.6	27.8	10.3	28.2	38.2
YOLOv3	Darknet-53	608	14	28.0	53.0	29.6	13.7	30.7	36.9
YOLOv3-SPP	Darknet-53	608	16	31.2	56.2	33.5	15.6	33.4	41.4

CenterMask: real-time anchor-free instance segmentation [21]
CenterMask-Lite	MobileNetV2-FPN	600x	27	25.5	—	—	9.2	27.1	36.3
CenterMask-Lite	VoVNetV-19-FPN	600x	20	31.2	—	—	14.9	33.0	41.2
CenterMask-Lite	VoVNetV-39-FPN	600x	12	35.7	—	—	17.8	38.6	48.5

EfficientDet: scalable and efficient object detection [22]
EfficientDet-D0	Efficient-B0	512	26	29.0	47.2	31.2	7.3	33.3	46.2
EfficientDet-D1	Efficient-B1	640	23	34.6	53.8	37.5	13.1	39.8	51.0
EfficientDet-D2	Efficient-B2	768	16	38.2	57.3	41.2	17.9	42.4	53.6
EfficientDet-D3	Efficient-B3	896	13	41.3	60.6	44.6	21.6	45.1	55.1

EAWNet: an Edge Attention-wise Convolutional Neural Network for real-time object detection
EAWNet	EAWNet	416	31	36.1	59.3	39.1	18.5	40.0	53.3
EAWNet	EAWNet	512	24	38.8	60.8	42.7	20.8	42.4	50.7
EAWNet	EAWNet	608	17	39.7	62.2	43.3	23.1	42.8	48.6

FPS, AP, AP50, AP75, APs, APm, and APl represent the frame per second, average precision, reach to 50% average precision, reach to 75%, average precision, and average precision for small-, medium-, and large-scale objects, respectively.