SAMGeo 2 uses high RAM for image with many box prompts

[l-bashaw]

Hi,

I'm trying to use SamGeo2 to segment a large tiff (~0.7GB) that contains 300 box prompts. I am getting RAM usage up to 120GB (and therefore have to run on CPU), and it does not seem too decrease if I lower the batch size. Is this because of the large number of box prompts?

FYI, i pass my box prompts as a path to a geojson file

[giswqs]

Is your GeoTIFF compressed? What's the uncompressed file size?

[l-bashaw]

Also, the output is what I desire. The problem is that inference uses 120GB RAM and takes 20 minutes. I know I'm using CPU so it will be slow, but I have to because of the RAM issue

[giswqs]

Try reducing the number of prompts. Note that width and height of each detected object has the same size of the orignal image. The more objects, the more VRAM it will use.

[l-bashaw]

I will do so. Your library is absolutely fantastic and I want to say a sincere thank you! Myself and many peers make use of your work and it makes our lives easier.

One more question, would switching the dtype from uint8 to one of the other options help?

[giswqs]

Thank you for the kind words. Glad to hear that it helps.

You can try converting dtype to unit8. Not sure if it will help much. The source code might have already converted image dtype to unit8 under the hood. I have not worked on the package for a while now. Forgot what I did extractly.

[l-bashaw]

Got it - I'll just split up the input prompts and feed them in batches.

Thank you again!