Auto-crop images of animals (image computation)

Nature Talk
43

Thanks @pisum , understood, as you had predicted simply changing to another model such as “Global - SpeciesNet - Google” did not improve the crops, makes sense, they are just for identifying, not for the bounding box detections.

I tested a bit more, but tweaking the confidence treshholds did not help, therefor I have reported 2 bugs on AddaxAI forum (it seems my posts become auto-flagged as spam there, and the system then hides/deletes them, but I am in contact with the the AddaxAI project maintainer Peter van Lunteren to restore them):

However, having tested AddaxAI (with MegaDetector under its hood) a bit more now, I notice that, while it generally works great for larger animals, but for ‘smaller’ arthropods/insects it often cuts of parts such as legs and antennae, which is problematic for me … Maybe I’ll create another post with regards to that on their forum in the future, or perhaps I’ll reach out to MegaDetector directly via e-mail as per https://github.com/agentmorris/MegaDetector/blob/main/megadetector-challenges.md On the other hand, I often also do not have auto-recognition of insects that take up a very large portion of the image (let’s say 70%), but where the body of the insect is not on the picture entirely, we could call them detail pictures I suppose.

I’ll do some further reading, it might be interesting to use a custom model to address some of those challenges.

Here is an attempted overview of all the tools that use MegaDetector to some extent: MegaDetector/megadetector.md at main · agentmorris/MegaDetector · GitHub

46

Wow! User @benfh has now created a tool to auto-crop for iNaturalist!

https://github.com/nebotron/iNaturalistReactNative

Amazing! @benfh

47

Hi @benfh,

I am reading through the GitHub, and I understand I need an Apple Developer Account. I checked out the “Apple Developer Program”, but it turns out this requires a USD99/year fee if you don’t have a recognised Non-Profit Organization, is that correct?

Is there not a possibility that you could publish it publicly on the App Store, such that others who would like to use it could simply download it on their mobile device?