search

msracver / Deep-Feature-Flow

Deep Feature Flow for Video Recognition
MIT License
1.3k stars 297 forks source link

Why is the last layer in resnet101 need dilation=6 ? #27

Open andyhahaha opened 7 years ago

andyhahaha commented 7 years ago

I read the R-FCN paper.

After the Resnet101 they append a randomly initialized 1024-d 1x1 convolution layer to reducing dimension. However, your implementation append the a 1024-d "3x3" dilation="6" convolution layer. The paper doesn't discuss this difference. I wondering if I use 1024-d 1x1 convolution layer like R-FCN, whether the result will be different or not?

Thx for your help !

liurongdev commented 5 years ago

I read the R-FCN paper.

After the Resnet101 they append a randomly initialized 1024-d 1x1 convolution layer to reducing dimension. However, your implementation append the a 1024-d "3x3" dilation="6" convolution layer. The paper doesn't discuss this difference. I wondering if I use 1024-d 1x1 convolution layer like R-FCN, whether the result will be different or not?

Thx for your help ! Can you add qq private chat? 1069919773