12 comments

[ 0.21 ms ] story [ 36.8 ms ] thread
I guess what I learned from this is google image search doesn’t use the same technology as google photos search. Photos is absolutely capable of “dog on a beach at night” and many terrifying things like finding people standing in front of small photographs of helicopters when I search for “helicopters”.

ETA: and the result from google photo search for “boy playing with toy helicopter” are perfect while that search on this new tool produces essentially random images: a lizard, a fishing pole, etc. I’m not sure the art of machine image segmentation and classification has been advanced here.

Thanks for the reply! The search results are a function of the images we have - we've only indexed 2M images so far. Also, the model prefers certain semantic patterns that we haven't automated yet. For example "a boy playing with a toy helicopter" gives more meaningful results. :)
Interesting. There are definitely two things going on here: understanding the photos, and understanding the question. Especially if I search for “ein junge mit einem spielzeughubschrauber” instead :-)

Honestly going up against the huge companies in ML seems like a big challenge.

"dog on the beach at night" is much scarier than "helicoper" when the image itself is relatively small.
"a picture of this website"

At least it's more or less consistent ;) great pics

a green stop sign
Just some feedback: search for “brown eyes baby” returns adults with blue, brown and green eyes.
(comment deleted)
My search for "endometriosis" was terrible
Flutter returns butterflies, flag returns flags and fluttering flag returns some kites.

But it's fast and clean.

I can confirm this works for pornography.
Pretty great, thanks, bookmarked!