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

Table 2

Ablation study on DCNv1, DCNv2, and our C-DCNets.

Method	Shorter side	Faster R-CNN					YOLOv4
	(600)				param	FLOP				param	FLOP
Baseline	Regular	39.4	60.8	42.4	51.30 M	100.1 G	41.7	63.5	44.7	26.8 M	146.4 G
Deformation	dconv@c3c5+	40.1	62.8	43.6	52.70 M	102.8 G	42.3	64.9	46.1	28.5 M	150.5 G
	dpool(DCNv1) [6]
Modulated deformation	mdconv@c3c5+	41.4	63.0	44.1	65.5 M	146.2 G	43.4	65.2	47.6	36.3 M	178.2 G
	Mdpool [7]
Curvature-driven	C-dconv@c3c5+	43.3	65.2	47.4	52.7 M	109.1 G	45.2	67.0	49.8	28.5 M	165.4 G
Deformation	Cdpool

The input images are of shorts side 600 pixels. And all settings are consistent with Table 1. The bold value means the best value of each item.