What Wildbook should this feature be in?
What would you like to see?
I’m helping Nikki Ladd (@nladd) and the Shark Experience team in New Zealand to submit their white shark (Carcharodon carcharias) encounters. It looks like machine learning has been implemented for dorsal fin IDs on the species, but we would like to incorporate other areas that are more useful for ID from underwater photos.
An example of the ID areas that Shark Experience (and other groups worldwide) use for white sharks is here:
How would this functionality help you?
Adding additional ID areas would improve functionality of Sharkbook for identifying white sharks from underwater photos.
Hey @simonjpierce , thanks for the request. I’m going to tag @jason because, as far as I know, these sort of ML requests, where we’re talking about all new part recognition, require a significant amount of data management and effort. I think it will be best for Jason to make determinations about this particular question.
Thanks so much!
Thanks @tanyastere! I know @jason was playing around with tiger sharks a while back, which have some overlap in ID areas, so I’m hoping that expanded capability for white sharks might have multi-species applicability.
Jason, the NZ white shark team have ~200 identified individuals. There’s overlap with the eastern Australian population – marker-tagged sharks from Australia routinely show up, and sat-tagged NZ sharks often head over there too.
I’m going to the White Sharks Global conference in South Australia in November, so it’ll be a good opportunity to talk to other teams about the potential net benefit of a regional or international photo-ID database.
Just giving this a nudge, as I’ve been talking to people at the conference and there’s definitely interest.
What would an appropriate training dataset look like for this?
(@jason – is this similar to what you were working on in tiger sharks?)
Sorry for the slow response. We do indeed already have above- and below-water dorsal fin matching in Sharkbook.ai using PIE v2 and the new, multispecies miewId. miewId was trained with a California white shark data set.
To match the other parts, we would ideally need at least 100 marked individuals with 4-10 photos of at least one side. But honestly, we don’t have a minimum or maximum number of individuals required in the new approach. Multispecies AI training is very new work, and we’re still trying to figure out the maximums and minimums.
First-pass training would likely just let the system learn from the whole body. We would blend this species in with others to get the real multispecies performance bump. Think you can get the data? We’d love to keep improving the support for white sharks.
Has there been any progress made on adding some new ID areas? I could provide some of our California White sharks dataset if we think adding these ID areas would be a fruitful venture. We have just over 2000 encounters with quite a few individuals re-appearing in our whole dataset going back to 1994. We primarily focus on the dorsal fins but I’m sure I could scrounge through our data and find videos we can pull screenshots from to get these new ID areas.
@Dmoran That would be great! If there are some known individuals that we could work with it would be excellent. There’s a lot of interest, but we haven’t received a reference dataset to train the system with yet.