Issue with MiewID match results

evgils · March 12, 2025, 9:52am

What Wildbook are you working in?
ACW

What is the entire URL out of the browser, exactly where the error occurred?
Bulk import:
https://africancarnivore.wildbook.org/import.jsp?taskId=db5f05b0-491f-4226-a23c-6f6a54354570

Can you describe what the issue is you’re experiencing?
Matches are not showing up in match results, while there are good photos in the system. Looks like algoritm is not including all photos that are available in the system to look for matches.
This results in users to create new individuals for photos, while that indiviual is already in the system. Which means we are creating duplicates!!

Can you provide steps on how to reproduce what you’re experiencing?
Go to bulk import 05-Mar-2025_Benny:
https://africancarnivore.wildbook.org/import.jsp?taskId=db5f05b0-491f-4226-a23c-6f6a54354570
We know this was a sighting of Kakumbi pack. We have good photos of this pack and IDs in the system.
Open the match results. No matches for Kakumbi dogs show up, while we have good photos in the system. Some photos do match with kakumbi dogs, others don’t.
The algoritm doesn’t seem to do a full search of all photos in the system.
For example the first encounter in the bulk import:

When I open the match result, it doesn’t show any kakumbi matches. I know this is individual EWD 1632. When I open EWD 1632 in ACW, I clearly see there are good photos of that side in the system which should match with the unIDed photo that I uploaded.

The weird thing is that some encounters do include the right photos in the matching results.
For example, the second encounter in the same bulk import of 05-Mar-2025_Benny.

This photo matches well with EWD1437 that is know in the system.

It is not a photo issue, since there are clearly good photos in the system that should have matched with EWD 1632.

Also 1632 is not the only example, there are multiple encounters in this bulk import that should match, but don’t show proper results in the match results.
All the encounters in the bulk import that include an ID resulted in proper matches, the other encounters didn’t, while good photos are available to match against.

I tested with multiple user accounts and tried to start another match, but the same issue continues.

I tried to figure out what the difference is between the photos that do and don’t show up in the matches. But they even come from the same bulk import.
For example this encounter for 1437 did show up in the match results:
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=656cb5c3-b374-4da0-b723-d6a31e50b9ba
While this encounter for 1632 did not show up in the match results:
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=8e14d2b4-09a2-402a-a203-68871ac61caa

When I look at the import details, they turn out to have been part of the same bulk import on 2024-09-27:
https://africancarnivore.wildbook.org/import.jsp?taskId=0dbe0c8e-1184-45e5-9504-4c5ba27a96bf

It would be very helpful if WildMe can look into this issue, because if we can’t rely on the algoritm to include all potential photos in the system, we would basically create a lot of duplicate individuals which would completely mess up our ID system.

Thank you!

If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report

evgils · March 12, 2025, 10:01am

In addition some photos that explain that it should have been an easy match for ACW.

Uploaded photo:

Photo of EWD 1632 in the system:

Both very clear photo that should have matched.

Strangely enough the encounters for the left side of this individual did show match results.

Anastasia · March 12, 2025, 7:47pm

Hi @evgils

Looking at your first example for this match result that should have ZCP_EWD_1632 as a match candidate, it looks like there’s no location ID on this encounter, only the country. Match candidates are determined by the taxonomy, viewpoint, and location ID. If you choose Zambia as your location ID when you edit the encounter and then restart your match, you should see an improvement in your match results.

I spot-checked other encounters in this import and none of those have the location ID, either. My recommendation is to update your spreadsheet to fill in the Encounter.locationID field, delete this import and upload it again with your corrected spreadsheet.

This does not appear to be related to the MiewID issue reported in Whiskerbook as I wasn’t able to replicate the same issue in ACW. All testing I’ve done has returned the correct match candidates.

evgils · March 13, 2025, 6:07am

Hi Anastasia,
Thank you for coming back to my post.

Last week, I already tried to re-run a match on Zambia location ID, but that didn’t yield any results.

I can try to upload the same batch with another template that specifies the ID already in the template. See if that helps. But I would think that the systems compensates for that when selecting the location ID for the identification process.

Kind regards,
Elke

evgils · March 13, 2025, 9:41am

Hi @Anastasia

The same issue keeps happening.
Hereby the details:
We did a new bulk import
https://africancarnivore.wildbook.org/import.jsp?taskId=a57e296a-41df-475e-b129-83fd2af7c620

ID location has been set

[image]

Match results:

Photos don’t show correct matches. We know it is dog EWD1842 which has good photos of left side in the system. But they don’t show up in match results.

[image]

Could you please check this import?

Kind regards,
Elke

Anastasia · March 13, 2025, 6:58pm

Hi @evgils

Thanks for following up. I resent this annotation to ID with Hotspotter also selected and the new match shows ZCP_EWD_1842 as a top-5 candidate in both results now: Wildbook for Carnivores

Can you let me know if manually restarting a match on one of the encounters in this import doesn’t show an improvement in match results? This information will help us understand next steps.

evgils · March 17, 2025, 12:36pm

Hi @Anastasia ,
I have tested.
The miewID algoritm didn’t results in any match

Neither the hotspotter. It doesn’t even show bad matches. There is nothing at all.

Would be interesting to hear from you how that’s possible.
Kind regards,
Elke

Anastasia · March 17, 2025, 10:10pm

Can you send me the link to this match page?

evgils · March 18, 2025, 6:50am

this is the link to the hotspotter match

Anastasia · March 18, 2025, 6:45pm

Thank you! Putting here for my reference that the above match is for ZCP_EWD_1632

Anastasia · March 18, 2025, 8:43pm

The reason that the right candidate is missing from the matches has to do with the IA class being assigned incorrectly. In previous encounters of 1632, the IA class is wild_dog_dark but for this encounter it’s wild_dog_standard. Different IA classes aren’t compared against each other.

We’re still looking into this, but it would explain why the match candidates for some of the encounters in your import aren’t showing obvious matches.

I’ll follow up when I have another update.

Anastasia · March 18, 2025, 10:40pm

Your response is unnecessarily adversarial, @PaulK. I’ve only shared what was relayed to me by the devs.

jason · March 19, 2025, 12:07am

@PaulK

While I agree with your historical interpretation of multiclass matching for wild dogs, that is not how the code is currently behaving on any Wildbook.

Anastasia escalated this issue to the dev team and has correctly summarized the root cause. This issue affects MiewID and HotSpotter equally.

Please allow us to complete our support of this issue on behalf of the originating user.

Thank you,
Jason

Anastasia · March 25, 2025, 9:14pm

Hey @evgils

I wanted to check in to let you know we’ll need more time to work on a solution for this.

When Hotspotter was the only algorithm in ACW, it worked well because it used a formulaic approach to pattern matching, regardless of which IA class was used. MiewID is very good at finding matches, but it does not tell us which features or characteristics it examines to suggest a match. It’s only looking at match candidates that share the same IA class, but it’s not yet clear if it has always done this or is doing so selectively.

We’ll follow up again when we have an update and have done additional testing.

evgils · March 26, 2025, 6:08am

Hi Anastasia,

Thank you for the update!
Hopefully you will get it fixed to match photos against all IA classes with MiewID as well. That would make it much more efficient for us to use ACW, otherwise we loose a lot of potential matches.

We will wait with uploads till a solution has been found.

Thank you.
Kind regards,
Elke van Gils