The mask generation time through point generation takes about 15 seconds??

[liujiadong369]

https://segment-anything.com/demo

I tested the point generation mask on the demo case on your official website, and it was almost seconds. Why did it take me 15 seconds to generate the mask on my server? Is there something wrong with it?

[swframe]

I took a look. The problem is that python doesn't have good multi-threading support so the masks can't be processed in parallel. One mask can take around 1 second to process and there can be a few hundred of them. I timed it but don't remember what were the specific calls but they didn't seem to be logic that I could rewrite to be faster. I did add the use of numba in a few places but overall the problem is that the processing needs to be run in parallel.

[e-tuanzi]

Have you ever paid attention to the image size? The larger the image, the slower the processing speed

[swframe]

I changed the code to only compute and save rle data and I resize the input image so the longest side is 1024. It now takes 2 to 4 seconds per image. I convert the rle output file to bitmaps and resize them separately (in java using multiple threads). It's that logic that can take 10 to 30 seconds per image in python code. The rle data also takes less disk space. My gpu mem usage also dropped to 30% of the available but running 2 sam processes didn't seem worth it