Description

hinerm

opened on Oct 15, 2024

Issue body actions

See imagej/napari-imagej#274 (comment)

Activity

1. hinerm commented on Oct 29, 2024

   

   hinerm

   on Oct 29, 2024

   MemberAuthor

   More actions

   Note that you can easily reproduce this in Fiji with a script like:

   #@ OpEnvironment ops
   #@ Img (label = "Input image:", autofill = false) img
   #@ Float (label="Sigma:", style="format:0.00", min=0.0, value=8.00) sigma
   #@output Img result
   
   # convert input ImgPlus to float32
   img = ops.op("convert.float32").input(img).apply()
   
   # apply gaussian blur
   img_g = ops.op("filter.gauss").input(img, sigma).apply()
   
   # subtract input ImgPlus from gaussian blurred image
   result  = ops.op("math.subtract").input(img, img_g).apply()

   and this sample data. It hits OOM if you run on the full image (352MB image, with 16GB of memory allocated in Fiji's memory and threads). If you split the channels first, it does not OOM.
2. hinerm

   added

   enhancement

   on Oct 29, 2024
3. hinerm

   changed the title [-]filter.gauss implementation scales poorly[/-] [+]filter.gauss implementation scales poorly with image size/dimensions[/+] on Oct 29, 2024
4. gselzer commented on Oct 30, 2024

   

   gselzer

   on Oct 30, 2024

   Member

   More actions

   I've yet to run this code, but I'd be interested to know where the OOM is happening - SciJava Ops more or less wraps gaussian convolutions from imglib2-algorithm, so this issue might best be moved upstream.

   Alternatively, maybe we should be wrapping different/better algorithms

Issue actions