I am new to Addax AI and just trying to work out the basics. I have a question regarding the use of species lists and how you think about them.
Now as the title says the project I work on is monitoring rats in 2 locations in Brazil other animals of importance to recognize would be opossums, cats and owls. Of course all the lists in Addax are mainly focused on wildlife and geographically the closest lists would be either the Colombian or Peruvian rainforests. But neither contain the animals of interest.
Do you think species like “spiny rats” and “ocelot” would work well enough to approximate the targets species?
I also saw SouthWest US containing “rodent”, “opossum” and “cat” so I guess a better match. But geographically farther away.
Otherwise do you think DeepFaun: “Micromammal” would be a better match?
And the last option is of course speciesnet, but I never managed to see a species list for that. The only thing anyone ever says is “2000 species” but I can´t ever see if rats are 1 of those 2000 species
Anyway I will of course test all of them and see what works, but I was just curious what more experienced people think about this or if there are any better options out there.
Nice project! There is indeed no species identification model for Brazil available. Even if there were, I would be cautious, since (1) Brazil is very large, and (2) your project likely uses a close-up camera setup that most models are not trained on. A close-up rat can sometimes look more like a bear than a rat.
Aside from the fact that some models would include your target animals, it is equally important that the other animals present in the ecosystem are also represented in the model. For example, you might have “rat” in a list, but if the model consistently mistakes opossums or cats for rats, then it will not be very useful. The background species matter almost as much as your target ones.
If you find that the target animals are identified reliably and not mismatched with other classes, then it does not really matter as long as you do not interpret the other classes. Likewise, if opossums are consistently classified as mongoose, for example, that can still be usable, because you know what is happening.
To get to this point, you need thorough testing. I would suggest creating a test set of a few hundred images that includes your target species and the other species found in your project, then running experiments with different models and settings.
However, SpeciesNet is currently not implemented fully in AddaxAI. It is still in Beta. That means that it runs, but there are limited settings, and the user interface is very bad. But apart from that, you can definitely try it out.
Awesome thanks for the help and pointers! I will definitely watch out for that. I unfortunately am not doing the practical camera setup in Brazil and don´t have much in terms of details on the setup, but the pictures look to maybe have a 5m range or so? Would you count that as close range?
