On 2020-11-16 17:08, Roy Smith wrote:
> Perhaps not exactly what was asked, but Google Images has the ability.
>  Go to images.google.com <http://images.google.com>, click the camera
> icon, and then you can paste a URL for the image.


This is perhaps the best suggestion so far. But it is limited.

Here I can input images from e-retail websites,
screenshots from Google Streetview, or just any
copyrighted photos and find similar photos.
If I add site:wikimedia.org to that search, I will find
free images on Wikimedia Commons. Good!

But the search is not so very useful. If I submit a
photo where much of the building is visible, I get
images of buildings in general, with or without lamps.
If I submit a photo where much blue sky background
is visible, I get light-blue images in general, with
or without buildings or lamps. There is a rough
visual similarity to the images, but not on a detailed
level and not by semantic content.

The search does not allow me, as I would wish, to refine
the search by indicating which images are better or
worse for my search.

It would be interesting to know what kind of data model
is behind the similarity function. My guess is that it
needs to be 100 to 1000 times larger.

Both in OCR (optical character recognition) and in
image similarity search, we now have so much data
that we should be able to organize a huge databank
of recognition data based on crowdsourcing. We'd
need some game, where millions of users answer
1. What does this image depict?
2. How well does this image depict XYZ?



--
   Lars Aronsson (lars@aronsson.se)
   Linköping, Sweden



_______________________________________________
Wikitech-l mailing list
Wikitech-l@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikitech-l