On Building an Instagram Street Art Dataset and Detection Model | Hacker News Reader

On Building an Instagram Street Art Dataset and Detection Model

blog.floydhub.com

49 pointsrememberlenny19 comments

Analyzing street art seems like something that will surprise artists when police figure out they can retroactively identify all the places someone vandalized (if they're ever caught once and tied to a particular style). I wonder whether artists will figure out how to confuse or otherwise prevent these kinds of automatic identification?

This is a great question that I definitely don't know the answer to yet.

There is a trope around graffiti artists that their biggest fans are the police. Often an artist will spend all night painting a mural, and then by the morning its already gone, but the paint "buffers" submit a photo to the police.

I definitely think there will be a cat-and-mouse synonymous with the adversarial neural networks examples for hiding weapons in everyday objects.

At the end of the day, most street art is commissioned or done in pre-approved areas, so I don't think it will be a cause of concern with larger artists.

That would be a neat foia project - request graffiti datasets.

Never thought of that, but this is brilliant.

Most large metropolitan areas already have "Vandal Police" who track serious vandals. Has been going on for decades, if they grab a big name (sometimes spanning multiple metros) they can pin them for hundreds of crimes resulting in six figure fines or large stints of jail time. These are usually people not painting murals though.

Just while riding my bike around I have started to remember the different scribbles I see sprayed on the walls. I have no doubt someone who is actively looking would be able to do a pretty good job given it seems these people spray the same thing every time.

This reminds me of an infamous graffiti artist in London, Tox https://en.m.wikipedia.org/wiki/Daniel_Halpin

Literally every single tube line/carriage/station has his tag, tox followed by the year, eg tox 16, repeated over and over again

If the police go by that, they are prone to being fooled by someone framing that artist: faking their identifiable elements (style or content), but perpetrating a lot more damage.

on the one hand, id think the police have enough money and time to build a detection system like this. on the other, we have people building surveillance system for free in their spare time

Hey all! I'm Lenny and the author of this post.

For anyone interested in street art and machine learning, please check out my email newsletter around these topics.

Link: https://www.publicart.io/

Speaking of Google Image search, Google actually did a pretty nice arts and culture project on this a while ago: https://artsandculture.google.com/project/street-art

I wonder if you could use their provided StreetView to generate a bunch of perspective views for the same artwork... or if computing a view from another angle could be a way to fuzz your existing training set to generate more examples.

You are one step ahead of me.

I want to try and map out street art in cities by using the landmark points in the background of images. Theres a few neat datasets that do landmark -> geodata translation, but as expected, its limited.

That being said, I can imagine in a very near future, it will be possible to train against google maps and get enough "landmarks" to geolocalize images.

God damnit Hacker News...as someone who recognizes the artwork in the header image because the artist is a good friend and was at my wedding several months back...

This is straight up surveillance that will cannibalize what it's observing.

They're quite creative, they're gonna dazzle your machine vision. Why build tools to hinder/distort/build paranoia into something you're interested in if you genuinely are?

Isn't that Instagram scraper (https://instaloader.github.io/) against their Terms of Use? Insta really locked down their API so I assume they won't look too kindly on this. (although it doesn't require login so they can't really punish accounts)

Zuckerberg is very happy to bend/break the rules and abuse his customers when it suits him, so screw him. These aren't instagram's images anyway - they're the images of the people who painted them, and the people who photographed them.

Its really horrible how locked down modern websites are. I wanted to write a script that would send me an email when a new event was listed on a facebook group but I saw that facebook has gone above and beyond to make sure the only way to see new events is to keep using their website.

Instagram has a really terrible web UI as well. When I want to share an image I see on it I have to open the inspector to get the image url because they blocked right click copy image.

It does require a login for stuff that you need a login for on the web interface (private accounts you follow and you also need to be authenticated to view anyone's stories)

It opens with the most silly value proposition anyone could think of! "What if you could pump all of the Instagram photos of Banksy's artwork into a program that could pinpoint where the next one's likely to be?"

Then it is a simple image classifier with one dimension (graffit, not graffit).

Hi, really neat project. I don't have much to add, other than it would be neat to see this for Buenos Aires street art. When I visited (granted this was about 10 years ago) I remember the amount of stencil graffiti was striking.

Reply on news.ycombinator.com