So "one unit"pictures on a neutral background - am I correct?
No multiple images and nothing in the background, because that means you have to edit the background out?
Yes that's about it.
Round 1 has 60 images of each and I've managed to get to the point where a paper cup with a lid is identified as an aluminium drinks can... which i understand because I deliberately ensured all images of paper cups did not have lids! That's round 2.
I guess all images of steel food cans ought to be unopened as well for round 1. Likewise maybe cardboard boxes complete (lids closed), and so on with all categories. I may have to have a big think.
Maybe then round 2 introduces lids on paper cups, opened boxes, opened food cans/tins and broken bottles?
There's a round 3 for refinements as well which i need to work on.
I have to fathom out confusion matrices next so I understand what it is doing and why!
So yes you are spot on for training images.
For test images, the item or items (all of the same category) can be partially obscured and amongst untidy backgrounds for some of the images (they are graded in levels of difficultly), provided they are the only category in that photo. So I can't Jane a glass jar behind a plastic lunch box (as I have found out) because it is homing in on the lunch box and calling it opaque plastic bottle which i guys is understandable. It is larger than the glass jar and much more dominant in the image. The fact that it is plastic it has picked up on, but it only has opaque plastic bottle and clear plastic bottle to choose from, so it's rather confused because it knows it is not a cardboard box, but that's the only box shaped item it has for identification purposes!
I think even I'm scratching my head at this point!
