Building an Internet Scale Meme Search Engine
findthatmeme.com
findthatmeme.com
One thing I'm curious about: how did you build your corpus of meme images and videos?
Have you tried something based on deep-learning that uses Transformers : https://github.com/roatienza/deep-text-recognition-benchmark (available weights are for tasks that seem similar to OCR so there is a good chance you can use it out of the box). With a good gpu it should process hundreds to thousands image per seconds, so you likely can build your index in less than a day. (Maybe you can even port it to your iphone stack :) )
https://github.com/microsoft/GenerativeImage2Text (You'll probably have to train on your custom dataset that you have constituted)
There are tons of other freely available solutions that you can get with a search for things with keywords like "image to text ocr" "transformers" "visual transformers"...
I’m around age 30, not 13, so similar to the article, my first instinct was also to create a database and OCR the image. But by total coincidence, yesterday I had a conversation with my 14 year old cousin on the topic of saving memes. Her response was along the lines of “yeah, everyone nowadays just saves the image to your iPhone photos, and then just search for it later from the photos app”.
Yeah. This whole article is literally already built into iOS UI, not just a hidden API. And kids all seem to know about this, apparently.
This article uses an example meme with the text “Sorry young man But the armband (red) stays on during solo raids”. I saved it in my iPhone photos app… and found it again through the search function in the photos app.
This is a solved problem already, by teenager standards.
I felt extremely old yesterday when I was talking to my cousin. And I felt extremely old today, reading this article. This is because looking back, the past few decades of CS cultural intuition have established that text are text, and images are images. Strings and bitmaps don’t mix.
This seems sort of obvious to anyone in tech, but I realized that from a clueless grandma perspective, not being able to search up text in photos wasn’t really obvious. Well, the roles are reversed now. Ordinary people now have access to software that treats text as a first class citizen in photos by default.
https://www.macrumors.com/how-to/copy-paste-text-from-photos...
I really want to emphasize how insane this situation is, because I think most tech people won’t realize what’s happening unless it’s pointed out.
If you’re a typical tech person, you probably look at this, go “oh, iOS Photos now OCRs every photo. Cool, that’s 2000 or 2010 era fancy tech, boring these days. And then a search engine on those strings, yeah cool, nothing too mind blowing”. The sheer boring-ness of this by tech people standards meant that this iOS UI change went under the radar.
That’s not true for non tech people.
The people who discovered this, put this to use immediately. You can search up anything from an image now. Old memes? Sure. Forgot the name of a restaurant you went to, but remember that you took a picture of the menu and the beef dish was amazing? Search up the word “beef” and it’s probably in there. Took a screenshot of an article, remember 1 or 2 words from it, but can’t find it on Google? Search for those 1-2 words you remember to find the screenshot, then use the phrases in the screenshot to find the article on google. Trying to find a picture of a cat you saved? Type in ”cat” and search for it. Yes, the photos app can do that too.
Screenshots are cheap and instant. Kids never delete them. It’s like how Gmail “archive” feature in 2005 revolutionized email because you never had to delete an email. Well, iCloud Photos “optimize storage” means that you can effectively store infinite screenshots.
There’s another UX revolution happening in terms of saving information. It’s just that photos became easily instantly searchable, and nobody seemed to have really noticed the implications this has on storing memories, and boosting recollection. This can possibly be the equivalent of “you’ll always have a calculator in your pocket” but equivalent analogy to memory techniques like spaced repetition.
Count me in. If you asked me about the OCR itself, I'd probably say "yeah, it's mostly been solved for a good decade for print books and articles, but it's unreliable enough". I somehow never considered OCR might have gone better - possibly because my main exposure was through badly OCRed book scans and a built-in OCR in some PDF reader I used at one point.
It definitely didn't occur to me that OCR works well enough on arbitrary images, and it's cheap enough compute-wise that you could do it locally in a casual fashion.
Nice thing you have there in the Apple garden. Over here in Android land, I have the opposite problem. You say:
> This can possibly be the equivalent of “you’ll always have a calculator in your pocket” but equivalent analogy to memory techniques like spaced repetition.
and all i can think of is how I recently became convinced that a Samsung flagship is losing my photos. There's been a couple cases over the past few months when I felt really damn sure I made a set of photos of something (e.g. remodeled kitchen), but when I checked on the phone, it turned out those photos don't exist, or there is maybe just one where I expected 5-10. They aren't in the gallery. They aren't in the filesystem. Poof, gone.
So either I'm getting senile in my 30s, or something is off with the way my phone stores photos. I did a web search for this the other day, there are relatively recent reports on-line complaining about the same thing, but no one has any evidence. I'm thinking about doing an experiment now (basically make extra photos every day and document them in a paper notebook, and check after half a year if the photos match the notes) - but the point of me sharing this is: I no longer trust new tech, smartphones in particular, to handle basics correctly. Much less do something advanced like reliable text search on images.
Note that Google photos not only OCRs, but it also does a visual search of objects, faces, scenery etc. and is extremely powerful.
I have Google Photos upload and backup both disabled.
But then, I'm pretty sure either Google or Samsung SMS app had a "feature" to automatically delete old messages (for a definition of "old" that was neither specified, nor configurable), and it defaulted to ON on my current phone, likely costing me significant chunk of my message archive (that I dutifully transferred over from the previous phone) before I accidentally found and disabled the switch.
So yeah, could be Google Photos deleting it. Or someone else. I don't trust Android as a platform anymore.
BTW. about this "delete old messages" "feature" - most likely this was implemented for performance reasons. But the thing is, you're unlikely to send or receive enough SMS in your whole life for it to take a noticeable amount of space. The irony here is, I do remember a case where the messaging app would become slow and laggy if you had enough texts stored on the phone - but that was solely because someone implemented the message list as a linked list, thus adding a O(N) multiplier to many GUI operations.
If you’re talking about kids without iPhones, then I don’t know, I’m assuming there’s probably some competitor apps on android now.
But I think you’re thinking too narrow. Don’t worry about a meme database. What about a searchable visual database of everything you’ve ever seen?
Though I share your feelings somewhat - I was completely surprised by what you wrote in the comment higher up about iPhone gallery search. I didn't realize this is possible in a reliable fashion, much less off-line and deployed in a mass market device.
The whole idea is that whenever you find a meme you like you save it, so all search is still local
I know the obvious response to the above will be “yes, really, teens do that. Pictures of everything”
But really? Everything?
You know that some teens don't even save some images? They store them on specific instagram accounts they make for a specific category. My cousin had an instagram account for close friends (5 people) where she only shared bad things that happened to her during the day. Another one for nice things, etc. All those memories were recorded in app and never saved in the device, only stored as stories on the account. Guess what? She was sad a while ago because somehow she lost access to one of this accounts and so to all those pictures/videos.
Also the fact that some teens save a lot of pics on their devices doesn't mean much in the grand scheme of things. In 2008 I had folders upon folders of images downloaded from the internet. Now i'm not even sure where they are, probably in some hard drive in some closet. You can be sure that if i remember any of those contents I won't dig up my hard drive, but google them (use a comprehensive global search engine). We have no idea where all this collected data from teens will be in 15 years, it's not unlikely that it will be lost or archived in hard to reach places and forgotten. I've stumbled some times into "meme dumps" where they upload all their memes to a service to free space on their device/icloud.
For sure teens use technology in ways that might be unexpected and counter-intuitive to us, but I don't think that invalidates in the slightly the need of a global search engine for memes. It's a good idea if i want to find a meme that I saw, a need that I don't think will disappear anytime and that's also not a millennial+ only problem.
If it works, it works. But it also speaks volumes about Apple's disregard/inexperience with exposing their stuff via the web - https://www.icloud.com/ being the prime example: half the stuff the phone apps can do are not available (cannot create a reminder with a due date...) and the things that are there are slow and buggy.
> Initial testing with the Postgres Full Text Search indexing functionality proved unusably slow at the scale of anything over a million images, even when allocated the appropriate hardware resources.
I can guarantee you that correctly setup PostgreSQL text search will be faster than ES with much, much less hardware resources needed, it's just a matter of correctly creating tsvector column and creating GIN index on it (and ofc asking right queries so it's actually used). I can help you out setting postgres schema up and debugging queries if you are interested, for testing purposes at least.
If you want to go beyond meme sites and possibly detect memes in the wild, common crawl might be something to start with.
People in the 21st century know a lot about the mistakes of the past century that led to much popular culture of the time being lost (especially terminally online people who've watched lots of Youtube documentaries about lost Dr. Who episodes and so on), so it surprises me how little we try and avoid those same mistakes with today's ephemeral pop culture in the form of memes. People like yourself who want to help make the internet's huge corpus of memes tractable are part of the solution in terms of meme archival and cultural memory.
(There's a good meme metadiscussion group on Discord, "The Philosopher's Meme," which you might be interested in joining. People there would be very keen to discuss what you've made.)
Memes are a core part of the Reddit experience, yet it's difficult to find something I know I saw before.
and specifically:
https://developer.apple.com/documentation/vision/vnrecognize...
He does mention running it on a macbook
Which would be slow expectedly
Low entry cost, recycling, eco-friendly, . . . a deck that writes itself.
At one of my previous workplaces, we discussed running the Z3 theorem prover on an iPhone cluster, because they run so much faster on A series processor than a desktop Intel machine.
My question is about the image distribution costs. All the memes on the site seem to be coming straight off an object storage, all that bandwidth consumption has got to add up(?). Some sort of a CDN might help depending on the search patterns.
Ideally there would be a best of both worlds where you could search memes by "characters" or "formats" in addition to text.
As feedback, it would be nice to search all memes from the homepage. The search on https://www.memeatlas.com/meme-templates.html also seems to be broken.
I'm so, so glad to see that I'm not the only person in the world with the same "problem". Well done, mandatory.
edit: holy crap you even index videos, nice
The framework is supported on macOS (even tvOS apparently) https://developer.apple.com/documentation/vision
If so, is there any risk in getting your account suspended or ip range banned somehow because of this, for example?
Now, after reading the article, I gave your search engine a try. I was looking for that futurama its a trap meme (pretty much pops up on any image search here https://www.google.com/search?q=futurama+its+a+trap)
The problem is, the search engine you built is now very text-heavy, which seems to be usually very unconnected to the actual meme. So, searching for "its a trap" did not yield the results I was actually hoping for, but made total sense looking at how the search was implemented.
Are you planning to implement an actual tagging of the content of some sorts? Maybe a clustering of similar objects (like iphone clusters similar peoples faces in the gallery) and then tag those clusters with keywords somehow?
I'd also like to figure out how to turn an image into a description of whats in it. My ML/tensorflow knowledge is very weak though, so I still have a lot to learn here.
is there really no open-source text recognition software that's on-par with or close to Apple's (presumably proprietary) implementation? the article mentions Tesseract. is that the current best open-source option?
Sometimes, I don’t know the exact words when looking for a meme, but once I see it, I know that’s the one.
IME CLIP embedding search can work strangely on memes as well, because it gets confused when images have words in them. Basically the same problem reported in the original CLIP paper where it thinks an apple and a piece of paper with "apple" written on it are the same thing.
I suppose that’s an old Intel MacBook? I’d be very surprised if the Vision framework performs better on a 2nd gen iPhone SE than even the first M1 MacBook Air.
Works great for me.
It's kind of like trying to come up with a good Google search phrase, based on how other people must have phrased something, but relying on knowledge of how you phrase things instead.
What was your average price per iPhone, if you don’t mind disclosing?
Fences worldwide will be overjoyed to hear of this novel application.
Do you have list of sources where memes are ingested from?
Would be nice to have some option to explore memes by category.
> Database Iceberg meme that shows different levels of database insanity
But possibly my google search results are tailored differently from yours.
I would love to replicate this setup for my own project....
I am thinking, load balanced, multi location redundant "iOS machines" with 3-4 phones in with power backup and internet dongle.
We could use something like zerotier/tailscale to get internet access from outside your local network
I once tried installing it on a recent Ubuntu. After messing with dependency hell and pip downloading half of the internet, when I finally invoked the CLI, it complained _again_ about a missing runtime dependency.
I called it quits. DL people are simply not interested in bundled static binaries.
Bonus: Index from a lot of sources to help track a meme's origin.
This type of thing is on my long list of "can somebody else please do this already".
Wish we could have used postgres but the tools were dictated rather than letting the requirements drive the tooling.
I believe I will actually use it a lot if you keep the site up.
Minor feedback for the blog post: It deserves a better meta description (for link previews). The first paragraph doesn't advertise how good the article is going to be.
Short TL;DR: It runs off my home server running a large vector database (opendistro): https://opendistro.github.io/for-elasticsearch-docs/docs/knn...
Maybe a dumb question, but could you use your data to train a new OCR model so you wouldn't have to rely on iOS?
I don't know much about ML/AI so maybe not feasible.
Quite a large bullet required, one with plenty of chewing left in it.
Only makes sense for small scale.
Unfortunate how there's no decent ocr library to self host, would be cheoaer than cloud costs.
Many places keep adding cloud services to their stacks until one day someone in the C-suite notices the AWS bill.