John Hunter writes:
I'm starting to think about adding image support and wanted to get
some input about what it should include and how it should be designed.
The ideas are still pretty nascent but here is where I was planning to
Create an image extension built around agg (not backend_agg). This
would be a free standing extension not tied to any of the backends
with the responsibility of loading image data from a variety of
sources into a pixel buffer, and resizing it to a desired pixel size
(dependent on the axes window) with customizable interpolation.
I guess I'm confused by terminology. What do you intend "backend"
to mean for images. A common interface for reading different
image formats? Speaking of which...
Inputs: what file formats should be supported?
* I can do PNG rather easily since I already had to interface agg
with png for save capabilities in backend_agg.
I guess I would argue for what you refer to below, that the
functionality to read image formats should be decoupled, at least
initially, from the plotting (display?) package. In fact, initially
it may make sense just to use PIL for that functionality alone until
we understand better what really needs to be integrated into the
display package. (The main drawback of PIL is that it doesn't support
either Numeric or numarray, and Lundt isn't inclined to support
either unless either is part of the Python Standard Library. It
may turn out that we could add it to PIL, or extract from PIL
the image file support component for our purposes. I suspect that
that stuff is pretty stable). But initially, reading images into
arrays seems like the most flexible and straightforward thing to
* As for raw pixel data, should we try to support
grayscale/luminance, rgb and rgba with the platform dependent byte
ordering problems, or leave it to the user to load these into a
numeric/numarray and init the image with that? Should we follow
PILs lead here and just provide a fromstring method with format
I haven't given this a great deal of thought, but again, arguing
for simplicity, that the array representations should be simple.
For example, nxm dim array implies luminance, nxmx3 implies
rgb, nxmx4 implies rgba. The I/O module always fills the arrays
in native byte order. I suppose that some thought should be given
to the default array type. One possibility is to use Float32 with
normalized values (1.0=max), but it is probably important to keep
integer values from some image formats (like png). Floats give
the greatest flexibility and independence from the display hardware,
if sometimes wasteful of memory. The second type to support would be
UInt8 (I admit I could be stupidly overlooking something).
These arrays are passed to matplotlib rendering methods
or functions and the dimensionality will tell the rendering engine
how to interpret it. The question is how much the engine needs to
know about the depth and representation of the display buffer
and how much of these details are handled by agg (or other backends)
* What raw types should be supported: 8 bit luminance, 16 bit
luminance, 8 bit rgb, 8bit rgba, 16 bit rgb or rgba?
Resizing: Generally the axes viewport and the image dimensions will
not agree. Several possible solutions - perhaps all need to be
* a custom axes creation func that fits the image when you just want
to view and draw onto single image (ie no multiple subplots).
* resize to fit, resize constrained aspect ratio, plot in current
axes and clip image outside axes viewlim
* with resizing, what pixel interpolation schemes are critical? agg
supports several: nearest neighbor, blinear, bicubic, spline,
Here again I would argue that the resizing functions could be separated
into a separate module until we understand better how they should
be integrated into the interface. So for now, require a user to
apply a resampling function to an image. Something like this might
be a good initial means of handling images.
im = readpng("mypicture.png") # returns a rgb array (nxmx3) unless alpha
# is part of png files (I'm that ignorant).
rebinned_im = bilinear(im, axisinfo...)
Then use rebinned_im for a pixel-to-pixel display in the plot canvas
(with appropriate offset and clipping). This isn't quite as convenient
as one step from file to display, but it should get us some flexible
functionality faster and doesn't restrict more integrated means of
displaying images. There are other approaches to decoupling that are
probably more object oriented.
I'll think more about this (and you can clarify more what you mean
as well if I'm confused about what you are saying).