On-Device Supermarket Product Recognition
ai.googleblog.com
ai.googleblog.com
Another thing I would like to do is compare products by price/quanity. Maybe I'm trying to get the most protein/$, or oz/$, etc. You could also hook it into some database to only show you ethically caught fish, organic projects, fair trade products, local products, etc.
I've thought about this kind of technology a lot. Cool to see that it's actually becoming possible.
Arguably, shops specifically design themselves to do things like this. When I was growing up, it used to drive my dad mad that the local supermarket would move almost literally everything around seemingly at random "to prevent shoplifting". He was a pilot, and usually had little time -- he liked to go into supermarket, grab x, pay, and leave. Suddenly, that stoped working, as finding x became increasingly difficult. It really annoyed him. For a long time the floor manager said that they moved their stock around to prevent shoplifting – until one day the store manager, when asked the same question, said it was "to promote footfall". Suddenly the real reason became clear.
>Another thing I would like to do is compare products by price/quanity. Maybe I'm trying to get the most protein/$, or oz/$, etc. You could also hook it into some database to only show you ethically caught fish, organic projects, fair trade products, local products, etc.
In the UK at least, all of this information is required to be printed on the label: product per unit money on the shop label and misery per unit product on the product's label.
Are you asking for a filter / search engine for that on top? Either way, I don't think your average supermarket would let that stay un-advertised for too long. Products at eyeball level pay a premium to be there.
Cant find anything about misery per unit, can you link to more information please? Interested to know how this is calculated.
I work on an app that does this. It's quite amazing how much technology has let us achieve in recent years.
[a] https://ai.googleblog.com/2019/11/introducing-next-generatio...
I'm wondering if 'on device' is an exaggeration. The post is saying 64 bytes per product, and a DB of millions of products, but the video is showing product details, prices, and a box image. I wonder if once it identifies the product it then loads metadata from ze cloud?
Flavour variants, special promotional packaging etc make training hard due to the sheer size of images that need to be in a training dataset to yield accurate results.
How is this achieved?
> OCR is executed on the ROI for each camera frame in order to extract additional information, such as packet size, product flavor variant, etc.
You could probably also do some very simple statistics like a colour histogram on the box, once you had high certainty of the product line.
Would cut the need to speak to shop staff, who in many places in larger cities don't have the language skills to know what you're looking for.
2. The supply chain for who made the hummus, who grew the chickpeas, were they sprayed with what. That basically does not exist, let alone in public.
Typically on-device computing is lauded from a privacy perspective when compared to cloud equivalents.
Shopping data is the most valuable. It is the end result of advertising.
What ad caused you to but that specific product? Can Google show a competitors ad to you?
You know this is a useful tool for visually impaired people, but I can't help but think that nobody should have to give up their privacy in order to have some level of equality. That is total horseshit IMO.