Only want to transfer Y (from YUV420)

[Hornochse007]

Hello everybody,

I think, this would be a very easy question, but I don't find the answer. Maybe anybody can help me?

I need only the Y-Plane from the buffer, but this really fast. So if I try to cut the np.array, the time will be to long. My idea was to use a new configuration.
Instead of the config {'format':'YUV420', 'size':(4056,3040), 'stride':4096, 'framesize':18677760}
I want a new format: e.g. {'format':'Y', 'size':(4056,3040), 'stride':4064, 'framesize':12354560} or the yuv420 and there only 3040 rows.

`
from picamera2 import Picamera2

pc2 = Picamera2()
cc = pc2.create_still_configuration(main={"size":(4056,3040), "format":"YUV420"}}
pc2.configure(cc)
pc2.start()
image = pc2.capture_array("main")
pc2.stop()
`

Maybe anybody has an idea for me?

Thanks a lot and stay healthy,
Hornochse007

[davidplowman]

Hi, can you just do

Y = image[:3040, :4056]

I think that will just give you a "view" of the data without copying it, so it should be fast. (Some operations want proper contiguous data, in which case you would have to copy it, it will depend what you want to do.)

[Hornochse007]

Y = image[:3040, :4056]
Y = np.ascontiguousarray(Y)
im = Image.fromarray(Y)

This skips the nearly the time from the conversion to PIL (max. 60ms)! Fantastic!! Thanks a lot!

Last Step: I would try to reduce the size of the readout of the buffer to only the Y-plane. Do you think i have a chance to change this parameters somehow? I tried to set the stride and framesize but couldn't do it. If I can reduce it the framesize to 2/3 of the YUV [so only Y - with the resolution, i think my time will shrink into 2/3 the time from buffer readout.

[davidplowman]

Yes, I'm afraid you can't choose your own stride, the system chooses the minimum valid one for you. Also there is no "Y only" format, and adding such a thing would be quite involved and require changing at least libcamera, but possibly also kernel drivers and maybe even firmware on the GPU (which is for most practical purposes impossible).

You could try using capture_buffer instead of capture_array and then using reshape. I don't know if that would be faster. I would certainly recommend using a size like (4032, 3024) if you can because that does make some things much faster.

[Hornochse007]

Ok, i will try and send a feedback.

[davidplowman]

Any feedback about a feature like that would come back to Raspberry Pi, so I think this is probably the right place to discuss it. However, given that it's a bit tricky and we have the YUV420 format, I don't think it would be regarded as a priority. Might there be other things we can do to help your code go faster? For example, you get can a low resolution image as well as the full resolution image and in some applications these can help.

[Hornochse007]

Sorry, i didn't get time to test, i will do it next week and give a feedback.

And where should i place the request for the Y-readout ? Here or in libcam?

I need only the high resolution to work with.
Thank you for your help and have a nice weekend!