Mask vs YOLOv3 performance

[RutenburgIG]

Hi folks,
I recently was looking to replace MaskRCNN with YOLOv3 since many resources report at least an x5 throughput bump with the second one. But I see following:

	MaskRCNN	YOLOv3	delta
Infer mean time (sec)	0.157	0.142	-10%
Infer median time (sec)	0.160	0.149	-7%

By infer time I get here model.forward() call. Configurations for both architectures are identical in the input size and batch size. So instead of x5 speed up, I have only 10% :( Does anyone have some ideas/clues about what's going on and how it could be?

[ElectronicElephant]

For YOLOv3, you should try a smaller input size.

To be exact, the default input size of Mask-RCNN is 1333 by 800, while for YOLOv3, it's only 608 by 608~

Definitely, YOLOv3 will become slower when input size is too large, due to its dense structure.