4 ms·

Google Image Search used to be perfect, around 2-4 years ago or so they swapped out the old school system which seemed to take some sort of fingerprint of the i

Google Image Search used to be perfect, around 2-4 years ago or so they swapped out the old school system which seemed to take some sort of fingerprint of the image and then tell what it was from the context of pages the image was found on. The current system seems to use machine learning to try and tell what the content of the image image is then just provide generic results for that term along with a similar color palette. Used to be able to find a movie screenshot on Tumblr, image search it and the name of the movie would come up. These days it'll go recognize the image is a woman on a street using ML, then show you results for "Woman" or "street" in the color palette of the image and if you're lucky you'll get a link to Pintrest too which also doesn't contain the context and just pushes you into a Pintrest onboarding flow. Feels like the Image Search team is more preoccupied with solving problems which are interesting to them with zero interest if the tool actually better or not for people who use it every day.

baybal2 6y ago
I think "machine learning" maybe a good approach for image recognition, but it is weak for image search, where visual hashes still reign supreme.
- hikarudo 6y ago
  The best descriptors are most certainly obtained with machine learning (embeddings).
  - wolco 6y ago
    The best are handmade.
sawaruna 6y ago
Yeah not sure what triggered the change but it was extremely noticeable.
- dublinben 6y ago
  The timeline matches up with their settlement in the Getty lawsuit, which many decried at the time as ruining the service.[0] It wouldn't be surprising if they chose to pivot away from the useful behaviour that got them sued, towards this inoffensive, ML-driven search that nobody asked for. [0] https://arstechnica.com/gadgets/2018/02/internet-rages-after-google-removes-view-image-button-bowing-to-getty/ https://arstechnica.com/gadgets/2018/02/internet-rages-after...
jjeaff 6y ago
Well, I think the new method is actually useful in a lot of cases and the old method has a totally different use case (i.e. finding more occurrences of the exact same image). The ML approach is good for finding similar images that aren't the exact same image. Which, honestly, seems like a much more common use case. The old fingerprinting method is pretty much just used for investigative purposes like finding out where an image originated from. Something I liked to use it for, but I doubt is a super common thing for your average user.
thatcat 6y ago
You can use tineye and it still uses the fingerprint method

baybal2 6y ago

I think "machine learning" maybe a good approach for image recognition, but it is weak for image search, where visual hashes still reign supreme.

hikarudo 6y ago
The best descriptors are most certainly obtained with machine learning (embeddings).
- wolco 6y ago
  The best are handmade.

sawaruna 6y ago

Yeah not sure what triggered the change but it was extremely noticeable.

dublinben 6y ago
The timeline matches up with their settlement in the Getty lawsuit, which many decried at the time as ruining the service.[0] It wouldn't be surprising if they chose to pivot away from the useful behaviour that got them sued, towards this inoffensive, ML-driven search that nobody asked for. [0] https://arstechnica.com/gadgets/2018/02/internet-rages-after-google-removes-view-image-button-bowing-to-getty/ https://arstechnica.com/gadgets/2018/02/internet-rages-after...

jjeaff 6y ago

Well, I think the new method is actually useful in a lot of cases and the old method has a totally different use case (i.e. finding more occurrences of the exact same image). The ML approach is good for finding similar images that aren't the exact same image. Which, honestly, seems like a much more common use case. The old fingerprinting method is pretty much just used for investigative purposes like finding out where an image originated from. Something I liked to use it for, but I doubt is a super common thing for your average user.

thatcat 6y ago

You can use tineye and it still uses the fingerprint method