On what GPU have you received FPS of 95/102?

[netanel-s]

First of all, thanks for the paper, I enjoyed reading it. In table 8 you compare FPS and COCO's AP with other CNNs, but I can't find the GPU in use. Could you please share on what GPU these FPS values were obtained? Is it Titan XP? If so, were the FPS values of the rest of the CNNs (YOLO, SSD, DSSD, etc) also computed on this GPU? (original values from corresponding papers are of Titan X Maxwell if I'm not wrong)

Thanks!

[TheMikeyR]

I'm also interested on which hardware you received these impressive results.

[zengarden]

U are right, I use TITAN XP GPU cause i don't have titan x. I think the speed will drop about 20% when adopting titan x.

[JingyuQian]

@zengarden I wonder, did you get the 95/102 results using your internal code? I trained this TensorFlow network, but I got surprisingly 2.3FPS on GTX 1080ti during inference. The input max size is 1333. I really hope that I could push FPS to 95/102.

[nguyeho7]

@JingyuQian The paper mentioned they used a different, lightweight backend network (Xception) to get 95 FPS. The network is not in this repo, as this repo uses resnet101.

[JingyuQian]

@nguyeho7 I see, thanks. But still, I think changing the structure to xception won't increase FPS by such a large margin.

[liuchang8am]

I assume the Xception like network used in the paper can't be shared ?