How to include -other- label when predict in AutoML Vision

Question

I have trained a model with 3 labels, but when using it to prediction phase, I uploaded an image that not belong to these 3 labels, it still detect that image is one of 3 labels. How to make it return result that the inputed image was not all 3 labels. If retrain a model with 4th label called 'other', I don't know how to prepare dataset for that label. Anybody have a suggesttion? Thanks for reading, sorry for my bad English

Answer 1

AutoML supports two types of models:

1. Multi-class - this is the default one - Multiclass classification makes the assumption that each sample is assigned to one and only one label. Keep in mind it assumes that always one label is correct and it will do its best to always return at least one label (actually it tries to make sure that sum of all prediction scores == 1).

2. Multi-label - this mode doesn't make any assumption, for each image there can be any number of labels detected - including 0.

You have two options:

1. When creating a dataset check the "Enable multi-label classification" checkbox
2. Add additional class 'other' or 'None_of_the_above' and add some example images to this class.

Options 1 and 2 can be combined.