Classification is suitable in the home automation context where a typical use case is to filter images captured by motion activated cameras, or determine some binary state (e.g. garrage door open/closed). The drawbacks of this technique are not a problem in this context, where we typically dont want to have to create massive datasets, and also we expect datasets to have a small inner class variation. The docs state 'However, this method works well when the training images (and the images to classify) have less variations. And in such case, only a few images (less than 10) are needed to train. For example, when an assembly-line produces a new type of defective part, you can use this method to teach the device what a defective part looks like on the fly.'
https://coral.withgoogle.com/docs/edgetpu/retrain-classification-ondevice/
Classification is suitable in the home automation context where a typical use case is to filter images captured by motion activated cameras, or determine some binary state (e.g. garrage door open/closed). The drawbacks of this technique are not a problem in this context, where we typically dont want to have to create massive datasets, and also we expect datasets to have a small inner class variation. The docs state 'However, this method works well when the training images (and the images to classify) have less variations. And in such case, only a few images (less than 10) are needed to train. For example, when an assembly-line produces a new type of defective part, you can use this method to teach the device what a defective part looks like on the fly.'