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.

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

The best descriptors are most certainly obtained with machine learning (embeddings).

The best are handmade.

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.

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

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...

You can use tineye and it still uses the fingerprint method

Yandex image search also uses facial recognition, which other image search engines have deliberately avoided.

As a result it's getting a lot of interest as a tool for investigative journalism. This tutorial by Bellingcat is really interesting: https://www.bellingcat.com/resources/how-tos/2019/12/26/guid...

Disclaimer: I only realize now that my comment is about images returned in image search results, not about reverse image search.

Bing definitely uses facial recognition: faces of celebrities are detected, there is a bounding box around the face, and their name and a link to Wikipedia. It does so for clothes as well.

Here is an example (fully automatic):

- the two faces: https://i.imgur.com/bfVNvW6.png

Here is another example, where clothes and items are automatically found in the image, and a bounding box is automatically defined:

- the necklace: https://i.imgur.com/vfhvY0T.jpg - the suit: https://i.imgur.com/AqS3qBV.jpg - the other suit: https://i.imgur.com/jTgRwpb.png

However, for the faces in this second example, I have to draw the box myself:

- the face: https://i.imgur.com/Pe1RLLm.jpg - the other face: https://i.imgur.com/LHOOwnW.jpg

So, what do you mean? What makes Yandex special with respect to facial recognition?

Google used OCR for routine indexing for at least a decade.

I found it how it was ridiculously good at finding car number plates, airplane numbers, and even yachts! https://www.google.com/search?q=рпв+2396&tbm=isch

In 2017 they did something to it, and it is now nowhere near as easy to find people's cars now.

> I found it how it was ridiculously good at finding car number plates, airplane numbers, and even yachts! https://www.google.com/search?q=рпв+2396&tbm=isch

Interesting! It finds this ad listing where there is a photo of a boat with exactly this registration number _in the background_ (the ad itself is for a RIB, or "Rigid Inflatable Boat" in front of it):

https://www.avito.ru/arhangelsk/vodnyy_transport/lodka_rib_s...

As far as I can tell the number doesn't appear anywhere in text form, so it must be OCR indeed.

I wonder, have they intentionally "spoiled" the indexing of car number plates to make it less creepy?

Importantly, it doesn't rank Pinterest highly, which is a major source of Google search spam

At this point it is pretty clear Pinterest have connections in Google which are pretty high up, allowing them to keep that behavior.

Is there concrete evidence of that or is that just the only thing that you think can explain it?

I can't dismiss that possibility, but I tend to think it's because, even though I don't see the appeal, with regular people, Pinterest is really popular and liked. In other words, the ranking seems wrong to me, but I think it might be because I am far from being a typical user.

A similar phenomenon happens with song lyrics on (regular) Google web search. If I search for "sting englishman in new york lyrics", what I want is the official lyrics from Sting's official web site (https://www.sting.com/discography/lyrics/128). Instead, I get pages of popular sites like genius.com, azlyrics.com, lyrics.com, metrolyrics.com, etc. I try to avoid these sites because their lyrics are often inaccurate and wrong (which defeats the purpose of looking up a line I'm not sure if I heard right), so the artist's official site or a fan site is vastly preferable. (I also want to support the artist. Maybe while I'm there I'll check for tour dates or merchandise.)

But I've mentioned this preference to some people before, and they surprised me by saying they like and prefer the big, well-known lyrics sites. To me, these are lowest common denominator junk bordering on spam, but to them it's what they are looking for. Point being, apparently most users aren't as particular about it as me and just want to go to familiar sites. Google may be ranking popular sites higher because that's what people actually want.

Google image search seemingly has become more about identifying your image and then searching for different images of the same thing. It’s really not a useful tool for identifying a particular image any more.

Yandex is my go to now.

Does Yandex have an api open to the public for image search? I’m working on a side project of an image recognition app that can recognize an image of a specific image a user uploads, search the web, and return the results. Exactly as shazaam works but in my case for images. Unsure if Yandex will be useful in this case. I looked into Google’s VisionAI image recognition platform but it seems that platform can only recognize what the object is (such as a shoe or cat) rather than giving me results of exact or similar images. If anyone here could point me in the right direction Id appreciate it.

I believe the Bing API can do this.

"Use smart identification of image content to recognize celebrities, find products, or search for related content."

https://azure.microsoft.com/en-us/services/cognitive-service...

Here are their APIs: https://tech.yandex.com/

At first glance, it doesn't look like Image Search is included.

VisionAI has similar image API as well. TinEye is probably the best option out there for your use case.

it also finds the image you are actually looking for, instead of the pinterest-of-a-free-but-not-free-stock-photo-spammer

it also shows the image full width, not in a frustrating slightly-bigger-than-thumbnail preview on a side pane

it also actually searches face images, not just tomatoes

also a lot less a prude than google

I was surprised when I discovered that Yandex's OCR is also better than Google's version:

https://translate.yandex.com/ocr

For example, comparing the results from pictures with Japanese or Chinese text, Yandex gives meaningful results, while Google often struggles.

Fascinating. I don't know if other people will get the same result but reverse image searching for [1] on Yandex gave literal CRTs in literal swamps, not just things in water or CRTs somewhere. Doing the same on Google gave neither CRTs nor swamps.

Regular image search is much nicer too but I wonder if that's just because it isn't polluted with products? Searching for galaxy on Google gave me all kinds of products while on Yandex is just gave me literal galaxies. Maybe just less spam because its less popular? Or maybe some of those Google results are actually ads?

1. https://im0-tub-com.yandex.net/i?id=a8f55cb11e37e5a6616b33bd...

Very cool results indeed. Clearly some object detection involved.

And Google give you "Internet meme" as a result.

Yandex image search is amazingly good, certainly for faces. Bing is also pretty good. Google is the worst out of these three.

If a picture of your face is hosted on the internet (for example on your blog) you can do the following experiment. Take a selfie (eg. a picture of your face that doesn't exist yet online) and upload it to Yandex. It will probably identify you.

Google image search is now terrible. Literally will just give you other examples of 'dogs'or 'man'

Yandex image search is frighteningly good, especially when it comes to facial recognition. Nothing comes close to it.

Another good example of this is this photoshoped image. Yandex give you both results, the background and the girl in the foreground.

NSFW-ish (girl in bikini) https://imgur.com/a/k8Ovoap

Some other handy tools:

- For finding the source of artworks, especially the anime/manga variety: saucenao.com

- For figuring out which anime a given screenshot is from: saucenao.com, trace.moe

- For a right-click shortcut for searching various image search engines: https://saucenao.com/tools/

I don't like that the "view full size image" is gone, or at least almost gone. I remember Google Images being very good in the early 2000's, but now it's really anti-user. It has become a chore to save images from Google, with a lot being Pinterest spam that doesn't fulfill my search.

This was the result of a lawsuit.. getty images if I remember correctly

Indeed. Getty is probably to blame for ruining Google Image Search.

https://arstechnica.com/gadgets/2018/02/internet-rages-after...

Yandex free services are underrated in general. I use them as a free email provider for all my programming related mailing lists and bugzilla accounts and I have had zero issues whatsoever.

I'm seeing decent results for both, then I tried Garbage: Yandex shows some random band I've never seen before, google shows garbage bins and landfills.

I won’t say it’s objectively better in all cases but Deepl provides better translations from what I’ve seen than Google Translate.

Are they still censoring anything that's considered offensive to Putin or Kremlin?

Probably, but that's only a problem if you live in Russia.