search

WilhelmusLab / ebseg

0 stars 0 forks source link

Lab to test FSD.py #27

Open mirestrepo opened 3 months ago

mirestrepo commented 3 months ago
tdivoll commented 3 months ago
danielmwatkins commented 3 months ago

The good: I was able to download the code, make a conda environment that just had jupyter and python 3.11, install the package, and run the sample workflow notebook.

The bad: The results for each processing level don't seem right. There are small floes in round 5 (gray) and large floes in round 4 (pink): image which overlap with floes identified in other rounds: image and don't really correspond to the amount of visible floes in the image. image What we should do is figure out whether the results from each round are actually being removed before the next round.

mirestrepo commented 3 months ago
  • [x] Add tests to compare pixel values

32 #34 #35

danielmwatkins commented 2 months ago

I pulled the updated ebseg library from GitHub this morning and re-ran it, and have started looking at the differences between the new results and Ellen's original imagery. What I'm seeing is the following:

Image from 2012-08-01 Screen Shot 2024-08-07 at 12 21 50 PM

Ellen's results Screen Shot 2024-08-07 at 12 21 40 PM

New results overlaid in shades of green and gray Screen Shot 2024-08-07 at 12 22 03 PM

cpaniaguam commented 2 months ago

I pulled the updated ebseg library from GitHub this morning and re-ran it, and have started looking at the differences between the new results and Ellen's original imagery. What I'm seeing is the following:

  • Most floe edges are the same between the two versions
  • In some cases, new floes are added by the new code, and in other cases, floes are missing in the new code
  • In cases where there are multiple floes in close proximity, floes are merged in the new code where they were separate in the old code.

@danielmwatkins The test suite is ensuring that the images are equal at the byte level. I added a couple more checks using a few other methods in #51 and they all check out. This might be something to investigate a bit more deeply.