Real-Time Human Ear Detection Based on the Joint of Yolo and RetinaFace

Table 2

A comparison of several object detection algorithm on the training set.

No.	Detector	Backbone	Input size	Training time	AP	AP	AP	AP	AP	AP
1	Faster R-CNN (2017)	ResNet-50		61,800	72.1	97.6	86.8	68.0	72.5	72.3
2	Mask R-CNN (2017)	ResNet-50		64,500	73.3	97.7	88.2	69.0	73.6	73.6
3	RetinaNet (2017)	ResNet-50		62,700	74.0	98.7	89.1	69.8	74.5	74.7
4	CornerNet (2018)	Hourglass-104		180,000	60.8	80.9	73.0	15.6	67.0	65.1
5	YOLOv3 (2018)	DarkNet-53		28,200	71.2	97.5	86.5	67.2	71.6	71.7
6	YOLACT (2019)	ResNet-50		38,400	71.3	97.5	87.3	66.3	71.6	72.6
7	Cascade R-CNN (2019)	ResNet-50		82,320	74.3	97.8	89.5	69.6	74.8	75.4
8	Dynamic R-CNN (2020)	ResNet-50		87,660	74.0	97.0	89.5	68.9	74.6	75.3