PyqtgraphViewer

[prybalko]

Image viewer based on Pyqtgraph lib

[prybalko]

There are couple issues:

• it's slow for unknown reason
• can't leave session properly

[tfarago]

@pavel-rybalko I took the liberty to try to use remote graphics which offloads the rendering to a separate process. However, there are two problems:

• the RemoteGraphicsView doesn't seem to integrate with the ImageView nicely (you cannot put there the whole view, only the image view without historgram)
• it is not working with gevent, I suspect there are too many event loops, because we have one from gevent and one from qt, maybe it is not the best idea to use both at the same time, on the other hand, I don't know how to circumvent it now.

[matze]

As far as I can see there is no problem integrating both event loops. But it seems that PyQtGraph implicitly starts the loop.

[prybalko]

I've added a gevent support for the viewer, but it's kinda dirty and still under development

[prybalko]

I've done some tests and measurements regarding performance of Pyqtgraph library.
I wrote a script, which creates a viewer and updates it in another thread / process as fast as it can. The results of performance are presented in the following table (clickable).

As you can see, the results are not really impressive and comparable with Matplatlib.

However, I found another library for visualization , which is OpenGL-based library called Vispy. And it works pretty fast (here is my script):

But this library is still under development and has some crucial problems with memory leakage (issue 1, issue 2)

Comparing all 3 libraries:

Vispy looks promising, but we have to wait till all bugs will be fixed.

[tfarago]

Nice summary, thanks. I would not go for Pyqtgraph and wait for vispy.

[prybalko]

ToDo:
latency test vs Matplotlib

[campagnola]

Sorry to interrupt; I got here by way of a related discussion on vispy. Your pyqtgraph test uses ImageView, which is not intended for video performance (the histogram computation and plotting is quite expensive). I recommend looking at examples/VideoSpeedTest.py for examples of the kind of framerates you can expect.

[matze]

Thanks for the pointer. I hope @pavel-rybalko could redo the test before leaving back?

[prybalko]

Here is an updated graph:

I didn't write my own script for testing, just trustfully used examples/VideoSpeedTest.py

Note: I limited FPS axis on the graph by 100

[campagnola]

That looks about right. Also note that the vispy framerates are probably limited by the GUI toolkit; their theoretical maxima should be much higher (so you will need much less CPU time to achieve the same framerate on vispy than other libraries).

[prybalko]

I wrote a script, which uses GraphicsView and update images in another process. It works fine, but we need to integrate it into RemoteGraphicsView (which implements multithreading) or to implement it like in my script (explicit multiprocessing). But already I don't have time for that though.