Curvature-Driven Deformable Convolutional Networks for End-To-End Object Detection

Table 1

Detection results on PASCAL VOC 2007 test set. The detectors are Faster R-CNN and YOLOv4.

Method	Shorter side (600)	Faster R-CNN			YOLOv4
		Baseline	Regular	69.7	78.6	62.5	70.5	79.8	63.4
Deformation	dconv@c3c5+	71.8	81.4	65.3	73.0	82.5	65.9
	dpool(DCNv1) [6]
Modulated deformation	mdconv@c3c5+	73.7	83.8	68.5	74.9	84.3	69.6
	Mdpool [7]
Curvature-driven	C-dconv@c3c5+	75.2	85.4	71.6	76.5	86.2	73.5
Deformation	Cdpool

The input images are of shorts side 600 pixels. In the setting column, “(m)dconv” and “(m)dpool” stand for (modulated) deformable convolution and (modulated) deformable RoIpooling, respectively. “C-dconv” and “C-dpool” stand for curvature-driven deformable convolution and curvature-driven deformable RoIpooling, And dconv@c3c5 stands for applying deformable conv layers at stages conv3conv5, C-dconv@c3c5″ stands for applying curvature-driven deformable conv layers at stages conv3conv5. The bold value means the best value of each item.