Multiple encounters from one photo

[Mcox]

What Wildbook are you working in?
Sealbook

What is the entire URL out of the browser, exactly where the error occurred?

• note the multiple encounter annotations + autodetect seal back as its body

https://seals.wildme.org/react/encounter?number=eeede11a-c5a8-4a71-9843-955d62c2d8bc

https://seals.wildme.org/occurrence.jsp?number=b014f636-10df-4da4-b322-4be21de65f3e

Can you describe what the issue is you’re experiencing?

Multiple encounters autodetected from one photo - Since using the new wildbook we have routinely been having issues with the autodetect creating multiple encounters for one known seal. Usually before submitting a request we have already broken it into one individual ( ex. RHS3) wild book will create multiple encounters off this one submission. - In the past we would go to all the duplicate encounters pages, untag and delete the encounter to narrow it down manually to one. This does seem to be occurring much more frequently and with more encounters per submission with the Sealbook,

Detecting water ripples as seal body - not sure how much we can do with this, but the autodetect feature is picking up the ripples as the seal body a lot, sometimes there is not even a head attached and its just a chunk of the ocean.

Duplicating photos - After we upload (lets say 4 photos) on the encounter page we will find that there are 6 photos of the individual (this did not happen in this encounter page but I will post it if this is not a user issue)

Not showing autodetected annotations - this may just be the speed at which things are happening so forgive me, but on some submissions the autodetect boxes do not show up in the encounter no matter time or refresh, the solution has been to annotate the photo - and when that is saved and we go back to the main page the autoannotations are there as well and we delete duplicate annotations.

Can you provide steps on how to reproduce what you’re experiencing?

We use sealbook classic so we can isolate our seals to the Howe Sound project as you suggested in the past, So

Submit> report an encounter (classic) > Fill out fields (select Howe Sound as project)> Submit

After the submission is processed we open the encounter and thats where our problems start. **[u]

[/u]
Lastly 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

[Anastasia]

Hi @Mcox

This looks a lot like what we saw in Internet of Turtles. What happened there was that the algorithm for green turtles had been trained on photos of turtles underwater taken by divers. Then researchers from another group started uploading green turtles photos taken from the surface where only their faces are visible above water (like your seals). This caused issues with detection because it was struggling to consistently annotate viewpoints it wasn’t previously trained on (underwater vs above water).

In this example, I see two head annotations and a body annotation. One head = one animal, so it created multiple encounters because it thinks it detected two heads. And it treated the water ripples as the body.

I’ll need to meet with the machine learning team to verify that’s what’s behind the questionable annotations and then I’ll follow up with next steps.

[Mcox]

Hi Anastasia thanks so much for being so quick with that!

I found some examples of the other issues for you.

Right after submitting that one one I went to upload another encounter and found an example of the duplication of photos (same pathway), as well as the hidden annotation boxes

Duplicating photos - submit 4 photos but in the Encounters image section there are 7 photos with 3/7 being duplicates of the 4 original submitted (

Hidden annotations - going through the unverified encounters this encounter shows 7 annotations, but inside the encounter page none are showing (As well as a creating a second encounter for this jan25 seal)

https://seals.wildme.org/react/encounter?number=1b065397-3de5-473b-8b30-88a59a0abfe2

[Anastasia]

Do you see that super tiny red pixel in the upper left corner of the image? There is a slight delay when annotations are generating and when their placement correctly shows up on the image. If I hover over it, I can see the hover text for the annotation’s IA class and viewpoint.

This is getting fixed in Wildbook 10.10. It’s supposed to catch up and display correctly after a minute or so but I’ve flagged it for them to review just to see if there’s any other weird behavior tucked in here.

[Mcox]

Thank you very much! Is it also duplicating the photos as a byproduct of the annotations?

[Anastasia]

Yes. If more than one head or body annotations exists in an image, Wildbook clones the encounter because each encounter represents one animal. So when detection makes mistakes, it creates more cleanup work until we can retrain the detector. Until I have an update, you can keep manually cleaning up annotations as you have been and we can eventually use your manually annotated encounters in our retraining efforts.