Data needed to launch MiewID on Amphibian & Reptile Wildbook

Hi ARW users!

You may have heard about our new MiewID multispecies algorithm from our Community announcements or from the Wild Me newsletter.

MiewID is much faster than Hotspotter, so you can see your match results sooner. Early versions of this model featuring only four species have been running on the African Carnivore Wildbook (ACW) and are already achieving more than a 10X improvement in ID job processing per day and even finding subtler matches missed by older algorithms.

We hope replace Hotspotter in ARW with MiewID, but we need your help in order to train the algorithm.

We’re looking for image sets of known individuals for the following species:

  • fire salamander adults
  • fire salamander juveniles
  • yellow-bellied toads

These should include only individuals that you’ve reviewed and are confident about their identities. The best way to do this is to provide us the URL to an ARW Individual Search that includes 100 or more individuals where that search result is filtered to data you believe has been curated accurately.

It is important to note that this advanced, multispecies AI can only be achieved through collaboration, and the field work of many different researchers contributed to improved accuracy and performance for all. Thank you to the many teams that directly or indirectly contributed to this new achievement in AI!

Let us know if you’re interested in training MiewID by responding to this post.

2 Likes

Tagging @Max.Muehlenhaupt and @David.Kupitz in case you’re interested in helping!

1 Like

Sounds exciting and I believe we have a suitable dataset of post-metamorphic fire salamander individuals. How many recaptures should each individual have at least?

@Max.Muehlenhaupt I don’t think there’s a minimum number of recaptures, but the more variety of viewpoints there are for an individual (whether that’s from multiple photos in a single encounter or multiple encounters) the better.

Ok, I will prepare a curated data set! However, each photo is +/- 90° angle from above of the individual. With different lighting conditions, resolution, focus etc. though.

1 Like

Ah, that’s right; variety of viewpoints may not matter as much in ARW since the original viewpoints Hotspotter was trained on were either from above or the side.

That sounds good! Let me know if you have other questions.

Hi Anastasia,
here’s the search result: Amphibian and Reptile Wildbook | Login
including 111 individuals that were encountered at least 5 times (up to 36 times). If we decrease the minimum (re-)encounter number, we would get even more individuals.
Let me know if you need anything else! Looking forward to the result.
Cheers,
Max

1 Like

With regards to fire salamander larvae and yellow-bellied toads, maybe @LauraSchulte and @MHederich can help?

1 Like

Hi Anastasia,
I am happy to help with the fire salamander larvae. Here we always have the same view point (right side) and very standardised pictures. And we don’t have super high recapture rates, I think 5 was the highest that we recpatured one individual. I will prepare something for you in the next days. It would be great to integrate the new algorithm!

1 Like

Thant sounds great! I think ARW and MantaMatcher are the exceptions for viewpoint variety because of the way their algorithms were trained. As long as the viewpoints are representative of your typical data sets, that will work.

Here a few links:

Is this enough? Please let me know if you need anything else!

1 Like

Thank you! I’ll share them with the machine learning team.