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

Table 3

Comparison results of different end-to-end detectors.

Model	GFLOPS/FPS	Params
YOLOv4	146.4/42	26.8	41.7	63.5	44.7
YOLOv4 with	165.4/40	28.5	45.2	67.0	49.8
C-dconv@c3 c5+Cdpool
YOLOv4 with C-dconv@c3 c5+Cdpool and the proposed loss function	168.5/40	28.5	46.0	69.2	51.3
DETR	80/19	37.4	38.8%	59.9	41.4
Deformable DETR	173/16	40	43.8%	65.2	48.5

The first row shows results YOLOv4 baseline. The second row shows YOLOv4 models with C-dconv@c3 c5+Cdpool. The third row shows results for YOLOv4 models with C-dconv@c3c5+Cdpool and the proposed loss function (11). The fourth row shows results for DETR model and the last row shows results for deformable DETR. Results are reported on the COCO 2017 validation set. The bold value means the best value of each item.