unsky / FPN

Feature Pyramid Networks for Object Detection
524 stars 263 forks source link

Do you consider "avoid aliasing" layer? #45

Closed twmht closed 6 years ago

twmht commented 6 years ago

According to the paper,

Finally, we append a 3×3 convolution on each merged map togenerate the final feature map, which is to reduce the alias-ing effect of upsampling.

But in the network you defined, there is no such layer.

Why?

unsky commented 6 years ago

in my code, it is 1*1