Cloud Vision Model Image Find No Similarity

So, im using Google Vision to detect watches, so i just follow the instructions, i create 2 labels (Komono and Casio) both are diferent watches.

I upload 10 photos in good resolution and with diferent angles.

So it looks like this:

So i just evaluate the algorithm, and when it finished a just use a photo with the watch on an arm (casio).

But oh dear it shows the "Komono" similarity?

How can i ensure that de model follows the similatiry more precise?

Regards in advance

Your training photos have similar background in each class. Try to variate the background of the training images, so it does not correlate with one class. Ideally the background should be of many types (e.g. street, grass, sky) and occur in both classes. This way the learning algorithm will optimize for watch-specific features instead of focusing on unrelated (background) information. To improve even more, add more photos to the training set.