Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Building an Internet Scale Meme Search Engine

findthatmeme.com

Post reply on HN
Print view View original

31–40 of 154 posts

Re: Building an Internet Scale Meme Search Engine

#31

Post by papito »

csande17 wrote: post #22

I wonder how the performance of Vision.framework on desktop Mac hardware compares to a cluster of phones. (The author mentions that it was "fairly slow", but it sounds like they were running an iOS app in the simulator and not a macOS app.)

Same thoughts. You can make that fly on a Mac Mini (provided it can be made to work close to the metal and not in an emulator)

Re: Building an Internet Scale Meme Search Engine

#32

Post by langsoul-com »

transitivebs wrote: post #5

My #1 recommendation for anyone thinking about the convoluted OCR solution: use a cheap OCR API and save yourself months of time / hassle / upkeep. Google's OCR API is a good place to start, but AWS has one too and dozens of others out there.

The author calculated the cost of 1 iPhone se was sub $50, which is 27k Google ocr images.

Only makes sense for small scale.

Unfortunate how there's no decent ocr library to self host, would be cheoaer than cloud costs.

Re: Building an Internet Scale Meme Search Engine

#33

Post by papito »

transitivebs wrote: post #5

My #1 recommendation for anyone thinking about the convoluted OCR solution: use a cheap OCR API and save yourself months of time / hassle / upkeep. Google's OCR API is a good place to start, but AWS has one too and dozens of others out there.

On the contrary, this sort of creative thinking is what's needed instead of automatically reaching for that shiny Cloud Toy. It's easy to get a proof-of-concept working, sure, but at scale, you start torching through cash.

Many places keep adding cloud services to their stacks until one day someone in the C-suite notices the AWS bill.

Re: Building an Internet Scale Meme Search Engine

#34

Post by formerly_proven »

yakubin wrote: post #9

The photo under Upgrading the iPhone OCR Service Into An OCR Cluster . In the future, data centres are going to host racks of iPhones.

Reminds me of imgix, who built their product on Apple libraries so ended up having racks of macs to run their service.

Re: Building an Internet Scale Meme Search Engine

#35

Post by philsnow »

I sat down literally last night and started sketching out the scratch-my-own-itch solution to more or less exactly this problem, because I too have meme-aphasia where I know there exists a meme that fits perfectly in a conversation, but I have about 5 seconds to find it before the moment passes.

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

Re: Building an Internet Scale Meme Search Engine

#39

Post by ducktective »

ipsum2 wrote: post #37

This is amazing. Out of curiosity, why not try deep learning OCR software instead of Tesseract? PaddleOCR is popular.

>PaddleOCR

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.

Post reply on HN