Did something change with the AI?

That photo (44509971) had a confusing background and emerging leaves, so it may have been a challenge for the AI. I cropped the photo to a portion of the frame, and it gave all woody plants as suggestions (I don’t know what it was so I don’t know how good they were).

Yes, that’s definitely a confusing photo! But I’ve had the same sorts of AI results with many observations in the past day or two, not just confusing photos, so I was just curious if something had changed.

1 Like

It suggests only plants for me in that observation, though all wrong, but better than birds.

1 Like

probably related to this:

with the new change, sometimes you’ll have to click the bottom of the dropdown list of taxon suggestions to include non-nearby suggestions, although, for me, the non-nearby suggestions for your specific example are further off base, i think.

i went back to one of my old observations where i had accepted a computer vision ID (mallard) for my initial ID (https://www.inaturalist.org/observations/69294668), and looking at the computer vision suggestions now, i don’t see that old (mallard) ID. it does seem like in the past, the computer vision would have easily made the correct suggestion for this observation, and now it does not:

UPDATE: see my next 2 posts for some additional info. what’s mentioned in this post is not actually the problem.

5 Likes

Wow, you are right! I am pretty surprised to see that, too. I believe the AI should have had no problem IDing those mallards, nearby or not.

1 Like

Oh, that’s interesting! I wonder why the difference.

I just looked at the new suggestions on some of my existing observations and they have clearly gotten worse. A good example is this observation:

https://www.inaturalist.org/observations/70105342

Previously Eastern Harvestman (the correct ID) was the top suggestion, but all of the other IDs listed were harvestmen. Now Eastern Harvestman is lowest suggestion, all all of the other suggested IDs are spiders. I’m noticing issues even on commonly observed birds like these Carolina Wrens (https://www.inaturalist.org/observations/68780046) or these Ospreys (https://www.inaturalist.org/observations/68739271).

I bet you’re right that it’s related to that change - thanks for digging that up!

1 Like

Would the suggestions be different when you drop the same photo in the upload form with no coordinates associated?

1 Like

with or without coordinates, it’s interesting that the upload page brings back better results:

i reuploaded a test observation with that same photo, and the observation page suggestions for the new test observation match the suggestions on the observation page for https://www.inaturalist.org/observations/69294668

in other words, the upload page hits a different API endpoint (/score_image) than the observation page does (/score_observations).

so it’s possible the bad suggestions on the observation page existed before the recent nearby / non-nearby change. usually, i make my identifications in the upload, not in the observation detail page…

8 Likes

Curiouser and curiouser. I was trying to add IDs to Unknown observations, not to my own observations, so I was only accessing the API from the observation page.

here’s a code comparison:

note the section highlighted on the left (/score_observations). i think what this means is that since the photo is now hosted at https://inaturalist-open-data.s3.amazonaws.com/photos/[xxx] instead of https://static.inaturalist.org/photos/[xxx], the computer vision is using the low-quality and croppoed “square” version of the photo for computer vision IDs on the observation page instead of the higher quality and uncropped “medium” version of the photo.

so the issue is actually related to the move of the photos from one host to another host, i think. (if that’s right, then computer vision code just needs to be updated to account for the new host location for the images.)

just to make sure, i downloaded the “square” version of my mallard observation and started an upload with it. the computer vision suggestions from the upload page are now similar to the suggestions from the observation page for my mallard observation:

10 Likes

Thanks for catching this! @pleary just deployed a fix, so try those problematic ones again.

@pisum is right about the underlying cause— because of moving some photos, it was using a tiny version. The image urls have been updated.

11 Likes

Thanks @pisum that was indeed the problem. Some poor coding on my part meant we expected photos to have a certain URL and recently many URLs have been changing. I’ve already made and deployed that fix so hopefully CV scores will be back to normal and unaffected by where they are hosted. This would have affected the CV results from the observation details pages of some observations whose primary photo has a CC or PD license, over the last two weeks

14 Likes

it looks good now. thanks to you all for fixing it so quickly.

7 Likes

Hooray! Back to adding IDs to Unknowns…

6 Likes

I’m noticing that many observations that used to get IDed by the CV as species that would be automatically marked captive/cultivated (and rightfully so) are no longer doing this because captive observations don’t seem to count for the “seen nearby” function. Could it be changed so that captive observations count for “seen nearby”?

Also, is it just me who is still experiencing very bad suggestions? I think the old way that the AI worked was much better.

1 Like

As an identifier of Australian grasshoppers I would make two points:

  1. AI stands for Artificial Ignorance.
  2. It could not be any worse.

The AI mostly suggests species from other continents. These are accepted by people with no expertise, not withdrawn and remain as mavericks.
Apparently it is not possible to only suggest species from the relevant continent. I would be happy to see AI turned off. It would save a lot of energy aside from the reduced frustration.

1 Like

AI now on default looks at seen nearby records, so you should see that, hopefully in next weeks new version will be better than it is now.
AI knows a lot of species and nobody forces you to use it, people who blindly agree will do that with any id, it can be a local species, still wrong, it’s a human error, not program. The more you observe taxa the better system recognises it.

1 Like

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.