Really bad results for image classification

I have several thousand images that I am trying to train for the ESP32 and the results are not good.

Project ID:

I am trying to train for different animals and have tried many different combinations on the training, none of which seem to give good results.

Any help would be appreciated.

And here is an example of one odd result -

Hi @delfin4,

Can you share your project ID and also check the accuracy for the unquantized model?


Hi, the project ID is 106160. I changed the images I was using but still see very odd results.

Try to check with two classes and also upload 100 images per class and see how is the accuracy then add the third class images and check it out. As long as for three classes 100 images per class will be enough.

Okay I will try this and thanks

Coming back to this issue, I am still seeing very odd classifications on images and wonder where the problem is.
If you look at the attached image, you will see that the image is of an elephant but the prediction is chimpanzee.

Hi @delfin4,

It looks like you are using MobileNetV1. Could you try MobileNetV2 96x96 0.1 to see if that works? V1 generally seems to have worse accuracy than V2.