-
Notifications
You must be signed in to change notification settings - Fork 18.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Understanding the deconvolution in FCN-32. #4952
Comments
Questions like these should be asked on the Caffe user mailing list or on another website, such as http://stackoverflow.com/ or http://stats.stackexchange.com/. |
From https://github.com/BVLC/caffe/blob/master/CONTRIBUTING.md:
|
If someone still confused with this, I found an answer: |
@warmspringwinds thank you very much! |
Hello,
I am trying to understand the design of the
FCN-32
model and especially the parameters of thedeconvolutional layer (convolution transposed).
Specifically, why the stride was chosen to be
32
and the kernel size64
.So, for example if the input image is of the size
768
by1024
.After the input is processed by all pooling layers we get
24
by32
subsampled predictions.Then the goal is basically to go from those subsampled predictions back to input image size.
Using the equation from the chapter
No zero padding, non-unit strides, transposed
from here and using stride 32 and kernel 64, I get output of size800
by1056
. Is it how it is actually done in the current implementation?I understand that after that we can just crop those to original input size.
My main question is: how did the authors come up with stride
32
and kernel64
parameters?I know that after all pooling layers the input gets downsampled by 32 but why the kernel size is
64
?Is it due to the fact that in the paper they initialize the filters to bilinear interpolation filter and wanted the kernel to capture the 4 closest points?
Sorry for posting it here. I just couldn't find the answer for the question in the paper or somewhere else.
The text was updated successfully, but these errors were encountered: