My two cents...

• Yes, we are planning integration with AddaxAI, stay tuned.
• Yes, you can run SpeciesNet locally. You can wait for AddaxAI integration, but if you want to try it sooner, you can run it using the instructions in the GitHub repo. We have tried to target the README for non-coders, and we appreciate your feedback on whether the instructions on the README actually hit that target, i.e. whether the instructions are clear, and whether everything works as advertised.
• Re: Indonesian species... we published a model card with a table of the number of images seen in training by country; Indonesia is fairly well-represented.  But I would add that whether the model saw a particular country in training isn't really predictive of how well it will do on new images from that country; the model only sees pixels.  We have seen great early results on images from Guatemala, for example, where zero images were seen in training; on the other hand, if you have a really odd camera angle in Canada, where we have lots of training data, the model will likely struggle, even though it's seen lots of data from that geography.
  
  But I would also add that whether an AI model can save you time isn't that related to whether it even saw your species in training, it's more related to whether it saw visually similar species in training, especially species that are visually similar to your common species.   If the model predicts a tapir from south America, and you're working in an area that only has Malayan tapirs, that's not really "wrong", we just need to do a simple mapping to call all the tapir predictions "Malayan tapir".  We look forward to working with the community to build up tools that make this kind of mapping easy.

That is super awesome!!! Looking forward to trying it out! And thank you so much for your explanation!

@dmorris , I noticed in the table that France is significantly underrepresented. If we know some people there, what’s the best way to connect them with you? Do you need additional data? Is there a regular training plan in place?

If we want to increase training coverage in new regions, the best thing you can do is encourage folks who are deploying camera traps in France to either (a) use Wildlife Insights for their work or (b) make their labeled images (not necessarily the location information, just labeled images) publicly available, e.g., on LILA.  With both of those approaches, the data becomes useful not just to us, but to anyone training similar models.

There are certain circumstances where we might take a special-case private delivery of data that's neither through the WI platform nor through LILA, but that doesn't scale well, plus in that case we would be the only ones who can use it, so it would have to be a pretty unique dataset to use that approach.

But also, as I mentioned above, it's tempting to focus on national boundaries at the atom of training representation, but it's not necessarily the case that more training data from France would make the model work better in France: a large training dataset with species that are visually similar to species that occur in France might end up being more useful to French users than a smaller dataset from France.

Which brings me to another thing you can do: encourage French users to try the model!  Or encourage them to email me, and I'll try the model for them. :)  No matter when/how we re-train, we need help from the community to identify areas where the model is less useful than we'd like it to be, and we need feedback in order to build good tools for mapping outputs in a way that makes it more useful (e.g. smart mappings of visually similar species given a known species list for a study area).  Adding new training data is always important, but it's not the most immediate need IMO, feedback from users and clever handling of the information we already have will get us pretty far.

Thanks, Dan! I’ll make sure to share this information as widely as possible.