Few-Shot Classifier

Quickly create a model using only a few images per item

The Few-Shot model can be thought of as a typical classifier that is supported by an underlying model with some predefined vectors so it has a base level of understanding. This is why Few-Shot can get away with using less collected data than a regular classifier. When you create a few shot model you will want to have at absolute minimum 2 images of each label. The maximum you should use for each label is 10. You are not prevented from using more but the speed of training will grow exponentially with more images per label.

The ideal set of data would be 4-8 unique images of an item. For instance, if you were trying to train it to recognize a pair of pliers you might want each of the following images:

• A pair of pliers laying closed on table

• A pair of pliers laying opened

• A pair of pliers in someone's hand

• A pair of pliers from a more top-down view

• A pair of pliers from a side-profile angle

Bonus points if you use as many different pliers as you have access to (different colored handles, mildly different size, etc).

If at any point you need help from the Passio team, please reach out to us at [e-mail]