Wild edibles app using image classification

Hello,

I created a wild edibles app using image classification in Peltarion. I used this dataset as “Edible” from Kaggle and added another class of ‘Toxic’ plants on my own.

Maybe it’s quite alright? But I only added around ‘200’ images to the toxic class vs ‘16,535’ images from the edibles in Kaggle.

Any tips on how I can improve my model? Or resources where I can get more wild toxic plant images?

Here is the link to my deployment, https://app.gcp-eu-west-1.platform.peltarion.com/148b70aa-a4e8-4439-a411-1fd5f28bdf04/https:%2F%2Fa.gcp-eu-west-1.platform.peltarion.com%2Fdeployment%2F148b70aa-a4e8-4439-a411-1fd5f28bdf04%2Fforward

Thankful for any feedback! :grinning:

Here is a blog I made previously about it, https://medium.com/@tosca.malm/building-your-ai-enabled-app-as-fast-as-lightning-in-peltarion-284345e3e35a

Tosca

This looks really interesting! I would definitely encourage you to consider evaluating your model with respect to recall instead of just general accuracy. Accuracy treats all kinds of mistakes equally (so, it doesn’t care if your model says a plant is poisonous and it isn’t vs. a plant is safe and it is actually poisonous). However, in reality, these mistakes are not equal! It is far more dangerous for your model to say a plant is not poisonous when it is (a “false negative”). Recall, or sensitivity, asks us: what percentage of the truly poisonous plants did your model predict correctly? This is far more important than if your model accidentally identified a harmless plant as poisonous! Here’s a useful reference:

1 Like