This is interesting.

I have always done RGB imaging with numpy using arrays of shape (height, width, 3). In fact, this is the form that PIL gives when calling np.asarray() on a PIL image.

It does seem more efficient to be able to do a[0],a[1],a[2] to get the R, G, and B channels respectively. This, obviously is not currently the case.

Would it be better for me to switch to this way of doing things  and/or work a patch for PIL so that the array is built in the form (3, height, width)?


On Tue, May 12, 2009 at 4:14 PM, David Warde-Farley <> wrote:

On 12-May-09, at 3:55 PM, Ryan May wrote:
> It's going to be faster to do it without the transpose.  Besides,
> for numpy,
> that imshow becomes:
>    imshow(b[0])
> Which, IMHO, looks better than Matlab.

You're right, that is better, odd how I never thought of doing it like
that. I've been stuck in my Matlab-esque world with dstack() as my
default mental model of how images/matrices ought to be stacked.

Am I right in thinking that b[0] is stored in a big contiguous block
of memory, thus making the read marginally faster than slicing on the

Numpy-discussion mailing list