New ML model to detect products on PRICE_TAG proofs

[raphodn]

Story

We already have a ML model to detect price_tag bounding boxes on PRICE_TAG proofs - see #511.
But it lacks context, such as the related product (above, or below the price tag).

Ideally train a new model that detects both the price_tag & the product (pairs).
Alternative is to train a separate model that only detects products.

[baslia]

I was thinking we can label images with 3 labels:

• product_and_price
• product
• price

Basically we will draw a bog bounding box on product + price, then 2 smaller bounding box on the product and the price.
If we respect that the product_and_price is always bigger than the area of product box + price box, the model would pick up on that logic.
Then on our side, with basic geometry, we can link back the triplet (product_and_price, product, price).
What do you think ?

[baslia]

Basically, the labelling task would look like this: