Chrome extension showing Computer Vision confidence

General
43

Yes, noticed this too… purple, teal, dark orange.

1 Like
44

I reopened this topic and moved the previous 3 posts here.

1 Like
45

Me too. And the colours were changing across the hours I was on iNat yesterday!

@sessilefinder is experimenting with ‘better’ colours? Maybe trialling better colour blind options?

Extension was last updated 11 Dec 2023.

1 Like
46

i’m not using the extension myself, but it looks like the raw scores are different than how i remember they were in the past. if i remember correctly:

  1. for species-level suggestions, there was a vision score, which was the raw score based on visual match only, and then there was also a combined score, which was the vision match score adjusted for a nearby factor. if i remember correctly, vision score and combined scores were would have been normalized / scaled so if you summed up each of these scores for all the suggestions, you would get 100.
  2. at a higher-than-species level, there was also a score that represented the summed species-level combined scores associated with the higher-than-species taxon.

what i’m seeing now is different from what i remember in two ways:

  1. the species level combined scores seem just be the visual score multiplied by the nearby factor, not scaled back up so that the total scores sum up to 100.
  2. the higher-than-species score is much greater than 100. it looks like these scores are 100x what i would expect.

so if you’re using the low-red-to-high-green scale in that extension, i would expect that when viewing things using expected nearby only, you would see:

  1. a lot of red where some of the high-confidence species-level scores would have been green in the past (since the combined scores don’t seem to be getting readjusted to total back up to 100).
  2. seemingly random hues for higher-than-species guesses (since they are going beyond the 0 to 100 range expected in the past).

image
image1875×966 178 KB

i see some changes to the computer vision from around the end of March (3 weeks ago), which i didn’t dig into very deeply, but which i assume would the source of the differences. i’m not sure what they were attempting to fix here exactly. so i can’t tell if these differences were intentional or not.

2 Likes
47

I’ve noticed that over the past week, the greens and yellows (ie high confidence) have entirely disappeared for me.

Fun. As others have said, I assume the underlying numerical score has changed. I’ll look into it as soon as I can and try to recalibrate the coloring logic.

2 Likes
48

if the changes were unintentional, i wonder if recalibrating the logic in your extension is the best path forward (unless it’s just a temporary workaround)?

1 Like
49

Indeed, now the color codes are different between the genus and the species guess. Seems that it works for the genus but species are in a strange blue color ?

image
image838×508 69.5 KB

50

i doubt anything has changed. the genus-level colors will be more or less random. if you look at 10 other observations, chances are you’ll come across a few that have odd-looking colors.

51

Since the new CV model was released this week I find the colors back to what they were before. I was rarely seeing green in agreement in the past model, but it seems to have reacted to the model update.

2 Likes
52

there does appear to have been a fix made (https://github.com/inaturalist/inatVisionAPI/commit/2510a98f5cfdb3aff2f297a09d92240bf0c6c3fe), but it seems to be separate from a new release of the model.

3 Likes
53

One more for the August updates list. Much appreciated that it can work as expected again!

54

My laziness has paid off!

1 Like