getBatch nearest neighbor interpolation about matconvnet-fcn HOT 4 OPEN

Hi, yes, it is. The main reason is that it is much faster.

I doubt that it would hinder the result. In fact, it may improve them as this is training data and the extra noise may help regularising. The amount of deformation is also fairly small. However, I never tested this properly (other than noticing that the performance of the trained networks is good), so it is possible that a different interpolation in training may slightly change the results.

On 13 Oct 2015, at 11:33, Amir Rosenfeld [email protected] wrote:

In get batch, the images seem to be re-sized using a nearest-neighbor interpolation,
look at line 77:
ims(oky,okx,:,si) = bsxfun(@minus https://github.com/minus, rgb(sy(oky),sx(okx),:), opts.rgbMean) ;

Is this on purpose ? Seems that it may hinder the results.

—
Reply to this email directly or view it on GitHub #1.

from matconvnet-fcn.

Well, I'm dealing with images which may be quite small to begin with, e.g, 60x60 pixels, so nearest neighbor interpolation may be quite detrimental in this case.

from matconvnet-fcn.

I've also noticed that while the getBatch function resizes all images to a common size, the fcnTest function does not.

from matconvnet-fcn.

Hi, yes, that’s because in training it is more efficient to stack multiple images together, which is only possible if the size is the same. For segmentation, this may be ok.

If your images are particularly small, you could e.g. resize them beforehand, or upsample them 2 times using bilinear interpolation before using the nearest neighbour warping. However, I am not sure that there would be an advantage in making the images smaller.

To be clear, with these networks, the input image size is arbitrary (within limits) as they are fully convolutional.

On 13 Oct 2015, at 11:40, Amir Rosenfeld [email protected] wrote:

I've also noticed that while the getBatch function resizes all images to a common size, the fcnTest function does not.

—
Reply to this email directly or view it on GitHub #1 (comment).

from matconvnet-fcn.