Hi, do you have any idea if resnet50 is supposed to be faster than VGG16?

Question

Hi, do you have any idea if resnet50 is supposed to be faster than VGG16?

yxchng opened this issue 7 years ago · 3 comments

Hi, do you have any idea why resnet50-faster-rcnn (~0.130s per pass) is so much slower than VGG16-faster-rcnn (0.060s per pass) when resnet50 itself is faster than VGG16?

I tried this repo: https://github.com/endernewton/tf-faster-rcnn and for this repo resnet-faster-rcnn is about the same speed as vgg16-faster-rcnn.

VGG16-faster-rcnn's speed is gotten by running Girshick's py-faster-rcnn repo.

Answer 1 · 2017-12-22T03:35:10.000Z

I think the complexity of the network after roipooling causes this phenomenon.
The VGG16-faster-rcnn only has two fc layers after roipooling, but the ResNet50-faster-rcnn has the stage-4 (3 bottleneck = 9 conv layers) after roipooling.

Answer 2 · 2017-12-22T06:51:10.000Z

Isn't the main complexity of VGG in the last two layers anyway?

Answer 3 · 2017-12-22T06:52:45.000Z

Also, if that is the case, why the speed is faster in the tf version? Which makes sense since resnet is supposed to be faster.