We are trying to improve our processing for adding images to our model, as Google recommends around 1k per label.
We had manually taken a ton of photos, over 10k across our labels and it’s very time consuming. We tried starting with a core set of images instead & use an post-image processor to replace the background so we can generate 500 images from a base of 50 originals.
Does this approach work? Is the same 50 images with different backgrounds going to have as much an impact as trying to take 1k photos per label?
I know there’s data we can try to use to verify, but we are hoping to hear from others on their experiences.