Allow "extract" outside original image bounds.

[danvk]

The problem was that y0-pad and x0-pad in psegutils.extract_masked could become negative. Chaos ensued.

This pads the output of extract with white pixels when you try to extract a region that goes outside the original image.

Fixes #18

[tmbdev]

Hi, thanks. That looks like a reasonable fix, but I'm concerned that extract now can return images that are of a different size than what the caller expects. I'm not sure whether that is going to break code somewhere. It's probably better to just return the exact size subimage that was requested but pad it.

Doing that directly is sort of fiddly, but it's actually easy to do with scipy.interpolation.affine_transform; you can give that negative offsets, desired output width and height (output_shape), and boundary conditions; just use an identity matrix for the matrix. order should be 0 or 1. Do you think you could change your patch to use affine_transform instead?

(I'd do it myself, but I'm busy bug fixing the clstm implementation right now.)

[danvk]

extract will always return images that are (y1-y0)x(x1-x0), which wasn't the case before.

I can take a crack at using scipy.interpolation.affine_transform.

[danvk]

I've reworked the code to use affine_transform and added a test image which reproduces the problem. Let me know if this looks right.

[danvk]

ping

[tmbdev]

Thanks; sorry, I was traveling. Thanks for looking into the patch. You still special-case interpolation.shift in that patch. The nice thing about affine_transform is that it doesn't need any special cases for any shifts. If you know what I mean, maybe you could simplify the patch to just use affine_transform for everything (mostly deleting a few lines of code), or I can do it soon.