Brush and Polygon Segmentation Loading Time

[anshul-dt]

Describe the bug
After Labeling Brush and Polygon, the loading time in tasks is huge, could take upto minutes sometimes.

To Reproduce
Steps to reproduce the behavior:

1. Label an image with Polygon or Brush segmentation, submit
2. Go to Tasks and try to load the image
3. You will observe that it takes a huge time to load (2-3 minutes sometimes)

Expected behavior
Should Load quickly

Environment (please complete the following information):

• OS: Linux Ubuntu 18.04
• Browser: Chrome, Firefox, Safari

[makseq]

@anshul-dt What storage do you use? Is it s3, gcs or local?

[anshul-dt]

local

[makseq]

What LS version do you use?

[anshul-dt]

0.7 and 0.8

[makseq]

I've checked it here: https://app.labelstud.io/api/project-switch?uuid=83ce5d2f-61b3-4409-ac30-863ae7f0d741
And polygons work fast enough. Brushes can be not too fast if there are a lot of strokes (RLE is a heavy format).
I think it will be easier to find the problem if you record video.

[tristanratz]

I had the same problem. Only thing that worked for me was downsampling the photos to a smaller shape

[makseq]

@tristanratz What image sizes do you have?

[tristanratz]

My images are sized 700px on the x-axis and dynamically on the y axis (according to the ratio)

[hlomzik]

Hi! That's sad, but currently brushes are not optimized, so their speed and size are badly depend on the image size (more specifically, number of pixels). So if it's possible to downsample your images, that's a good temporary solution.
I'm working on optimization, which increase speed significantly, will be released soon.
Sorry for inconvenience.

[iakremnev]

Hi, I've recorded some footage of this problem. I noticed following issues:

1. Brush annotated tasks take 10-15 seconds to load for 1920x1080 images. Even the project page loading is slow.
2. Browser consumes much of the memory, not the backend. And it easily takes ~6 Gib, which is too much for a dataset of 17 images.
3. Single CPU is used for forming and parsing RLEs. I understand this format is essentially sequential, but at least could be implemented in a separate process so that browser doesn't consume all the resources waiting for the encoding.

output2.mp4

[smoreface]

This should be fixed with version 1.3.0 release, please try it out!

[vchaparro]

Hi,

I'm using version 1.7.2, and it seems that the issue has not been resolved. I am using images of approximately 1500x1500 (around 5 MB) with segmentation polygons. When attempting to open these pre-labeled images, which have a relatively high number of polygons per image (it's about buildings footprint segmentation), it takes too long to open. Additionally, the user experience suffers when transitioning from one image to the next, zooming, and so on. The client-side web browser begins to consume high CPU resources.