WILDLABS logoWildlabs.net : The conservation technology network

WILDLABS logoWildlabs.net : The conservation technology network

discussion / Camera Traps  / 27 March 2025

Graphic interface for wildlife species ID of camera trap data?

Groups

There seems to be lots of machine learning options to identify objects and wildlife species in R or Python.  However, if you are R or Python "challenged" like me, it seems there are few options.  I’m currently using Megadetector through the graphical interface provided by EcoAssist.  EcoAssist via Megadetector works great for object detection.  However, we would like to use machine learning to identify PNW wildlife species and want to be able to train models.  Are there any existing machine learning - graphic interface for species identification?



Dan Morris
@dmorris
I help conservation scientists spend less time on boring stuff.
Inventory_key_contact
WILDLABS Event Speaker
WILDLABS Research Participant

27 March 2025 7:45pm

Training remains somewhat more of a hassle than inference, but thanks to a WILDLABS grant, our friendly neighborhood machine learning folks at DrivenData are working to narrow that gap:

https://wildlabs.net/discussion/wildlabs-awards-2024-no-code-custom-ai-camera-trap-species-classification

I know that's just a post about a thing they are going to build, but I wouldn't have posted that link if I weren't 100.0000% confident that it was not vaporware, so, stay tuned.

Also, I know I'm super-biased, so take this with a grain of salt, but in my experience the recent release of our SpeciesNet model changes the equation significantly re: when it's worth training your own model.  SpeciesNet saw tons of PNW data in training, and I've run it on a gazillion images from the PNW since its release, and it works quite well.  If it doesn't work as well as you'd like out of the box, consider doing some postprocessing as an alternative to training your own model.  It can be tempting to compare a model that doesn't exist yet to one that does, and assume that the former will be perfect, but no model is perfect, and even if the custom model would be a little better, the time it takes to train and maintain a custom model may be more than the time you would save thanks to the delta in accuracy.

OK, taking off my super-biased hat now, YMMV.