A Small Object Detection Algorithm Based on Modulated Deformable Convolution and Large Kernel Convolution

Table 3

Comparisons of the speed and accuracy of different object detectors.

Methods	mAP0.5 (%)	mAP0.95 (%)	Parameters (M)	GFLOPs
YOLOX	83.6	54.2	8.94	26.64
YOLOv3 [3]	70.8	33.6	61.54	65.54
YOLOv4 [1]	75.2	38.4	64.36	60.33
EfficientDet-d0 [18]	54.4	22.7	3.83	4.61
RetinaNet [7]	85.1	47.8	36.39	146.00
Faster R-CNN [24]	67.5	34.0	28.31	939.45
CenterNet [26]	87.9	51.6	32.67	109.34
SSD [19]	65.0	30.2	24.01	61.11
BGD-YOLOX (ours)	88.3	56.7	21.55	63.93