Tflite model labels missing the 'none' classification

[herbbetz]

Many Thanks for the nice model, but can it be, that the 400 labels have no classification representing "no bird"? So a picture containing only "me while eating a pizza" gets this result: Abyssinian Ground Hornbill (33.08%).

Greetings from Germany,
Herbert Betz
herber7be7z@gmail.com

[rprkh]

Hi @herbbetz,
Thank you for your question. Yes, that is correct. The 400 labels do not include a “no bird” class, which is why the model still predicts a bird species (in your case, Hornbill) even when no bird is present. The app assumes that the input image contains a bird. There are 2 possible ways to address this:

• You can add new images representing the "no bird" class and train the model with 401 labels.
• You can set a confidence threshold (say 40%) to reject low confidence predictions. In this case you would no longer get an incorrect prediction, since 33.08% falls below the set threshold.

Please feel free to open a PR. I think the confidence threshold fix might be a simple and easy solution to this issue.