search

ZilinGao / Global-Second-order-Pooling-Convolutional-Networks

Global Second-order Pooling Convolutional Networks (cvpr2019 GSoP)
MIT License
129 stars 29 forks source link

Is the second-order block suitable for ResNet18 ResNet34 network? #9

Closed wangxinyue11 closed 4 years ago

wangxinyue11 commented 4 years ago

I see that the second-order block is not used in the resnet network with a lower number of layers in your source code. Is it not good? Why is it not good?

ZilinGao commented 4 years ago

@wangxinyue11 Hello, Xinyue. Sorry for the late reply.

We conducted the experiments on ResNet18 (on ImageNet-1/4K). It also superiors to the baseline. And we did not evaluate GSoP performance on ResNet-34.

The results on ImageNet-1K shown in paper are all based on ResNet-50, so we only release the ResNet-50 backbone. If you are interested in GSoP performance on other depth of ResNet, you could deploy the network by changing the ResNet block number in each stage easily.

Thanks for your attention! :)