How to get the conv5 features with pre-trained ImageNet model for images with different width and height?

[zyzhong]

I want to calculate the pooled conv5 features for images 480x320. But the prototxt file seems always have the same width and height. I'm not sure in prototxt file which is the width.
Is the following prototxt file right?
I use the model trained in ImageNet2012 so the image mean is 256x256, how to get a reasonable mean?
And one more question, if I use this prototxt I get 14x9x256 features, does the pooled conv5 features map preserves spatial layout of the images?

input: "data"
input_dim: 1 //batch size
input_dim: 3 //channel
input_dim: 320 //height?
input_dim:480 //width?
layers{
...

[sguada]

That's a valid prototxt, for speed if you have enough memory you may want
to process more that one image at the time, by increasing batch size.

Yes caffe preserves spatial layout, but remember that you need to switch
dimensions W and H of the features due to row vs column order of Matlab.

Sergio

2014-09-09 7:12 GMT-07:00 zyzhong notifications@github.com:

I want to calculate the pooled conv5 features for images 480x320. But the
prototxt file seems always have the same width and height. I'm not sure in
prototxt file which is the width.
Is the following prototxt file right?
And one more question, if I use this prototxt I get 14x9x256 features,
does the pooled conv5 features map preserves spatial layout of the images?

input: "data"
input_dim: 1 //batch size
input_dim: 3 //channel
input_dim: 320 //height?
input_dim:480 //width?
layers{
...

—
Reply to this email directly or view it on GitHub
#1056.

[zyzhong]

Thanks a lot.
The mex function caffe() in matlab returns a 14x9x256 features map, do I need to switch
dimensions W and H? How to do it?
I use the model trained in ImageNet2012 so the image mean is 256x256, how to get a reasonable mean?

[shelhamer]

The input blob dimensions are batch, channel, height, width so your prototxt is right.

Note for non-fixed input sizes you need to make the model fully-convolutional: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/net_surgery.ipynb

For the mean the simplest approach is to reduce it to channels only i.e. average over height and width. That essentially works just as well and is simple.

Please ask usage questions on caffe-users. Issues are for development discussion. Thanks!

[shelhamer]

Fully-convolutional models plus @longjon's #594 in the latest release completely address feature map extraction for any input size.