Hi,
I've just picked up the new iRecord butterfly app version which now includes a butterfly suggestion feature.
This works well and I like the way it provides probability scores.
Would anyone be able to share some of the background on this feature - for example is there an ML model loaded onto the phone, will this be updated over time, how many images was it trained on, etc.?
Also, would it be possible to consider a workflow within the app (or a separate app) to enable 'beginners' to ID butterflies without first having to enter a species suggestion? I have friends that I am trying to interest in butterfly ID and an app for ID (that also submits a record) would be a much easier proposition than a recording app that also has an ID function tucked away after you have chosen species and uploaded photos.
I think there's a massive opportunity to collect many many more records in this way, and using photos will help remove misidentification. Even records of easily ID'd species are useful to improve datasets. A 'butterfly ID' app is one I can easily get friends to install and use to take photos of butterflies. A 'recording app for butterflies that has an ID function on this sub-menu' isn't an easy conversation.
Thank you.
Thanks for the information
Thanks for the information - I should have thought of just uploading an image before choosing the species!
I agree there's massive potential with this approach, especially in relation to citizen science and encouraging data collection / submission.
Are you able to share any more details on the external ML service?
I setup an AWS Rekognition Custom label model a while ago, but at $4 per hour it wasn't feasible to run it 24x7 for a personal project.
Are you using a public cloud hosted service, or something else? I'd be very interested to know if there's a more cost effective approach than AWS. I briefly looked at using on-device processing but that's much harder to do.
Your model seems pretty accurate from my limited use so far - how many images was the model trained on?
Thanks for taking the time to share your knowledge in this area.
Thanks for the suggestion. We are still at the early stages of integrating image classification support for ID, but also recognise the great potential. We want to understand how well the models work before making the ID function a more overt part of the workflow.