Hacker News Reader

Comment by BoorishBears | Hacker News Reader

This looks like such a cool idea, but already you see that the problems presented seem to be seriously underestimating how complicated what they're trying to do is. For example, detecting specific text in an image? 10 minutes in OpenCV. Detecting any text at all in any format in an image? I don't even know where to start. Maybe 10 minutes in OpenCV if they constrain the kind of text, otherwise ML? It feels like an unboundedly complicated problem.

View on HN

It's also not clear how it will be used later. It could be used to crop/remove watermarks or copyright messages, and I wouldn't develop it if that's its purpose.

Captcha solving is expensive for spammers.

are you sure? What's your source?

I can assure you it is not. It's about $0.001 per solved captcha with 97% accuracy rate and refunds on failed ones, from the service provider I know of.

I've even written one, it's not complicated and not expensive.

Can confirm, it was very cheap 5+ years ago when I was in the business of needing them solved en masse. I doubt it's gotten more expensive.

It is still the same rate. 2.2 ish for new recaptcha, too.

Yeah I was able to beat that "Robot check" checkbox by Google using some fuzzy bezier curves and random mouse movements/scrolling to simulate a real user's behavior.

Unfortunately Google has some other metrics to detect bots (such as no cookies showing they've seen you before, etc), so it wasn't 100% reliable.

you don't need 100% reliability, that's the good side :)

I have to assume that this is exactly what this would be for. 9gag or some similar site stealing content en masse from elsewhere and wanting to remove any sort of identifying marks.

Also my first assumption. When I read that task, being a retired professional photographer who made a living for 15 years with camera, my blood pressure rose.

Google's Vision API extracts text in any language or format from images: https://cloud.google.com/vision/

There are a bunch of SaaS opportunities like this that could be made just creating wrappers for various cloud apis and selling to people who don't know they exist.

Yep, this. Basically leveraging the knowledge difference between developers on the edge of what's available with Cloud ML API services, versus people with business problems that are not aware of same. Throw in some customization and nice marketing / design for niche use cases, and you've maybe got a business.

I can't speak for the site, but I think the idea is that you sort of get this continuous funding to keep working on and improving your solution. (Makes it kind of different from bug-bounty type approaches, which I like.)

I think it'd be nice to gauge the interest of such a project. There is one person pledging right now, but if more do and the demand clearly seems to exist - maybe it's worth investing in.

there is a pretty substantial gap between '$75 for at least one month' and 'something that will make enough before someone else does it too to make it worth the time'

There is already an API available to do this today. The Microsoft Cognitive Services computer vision API detects words, phrases, and lines and included bounding box information indicating where the text was found. See https://azure.microsoft.com/en-us/services/cognitive-service....

OCR excels at picking out digitally generated text, but it's not so great at "real world" text unless you help it by zooming in.

For example, it can't tell there's text in this clear image of a VIN:

http://1eask1khnn22hm3njx2dmz1a.wpengine.netdna-cdn.com/wp-c...

But I bet if you took a closer picture (like you would if you were using an AR app) it'd work

In terms of ML you could probably load a pre-trained image classification convolutional model and fine-tune it a bit and get quite a reliable if not superbly exact "here is some text" classifier without too much effort. Perhaps not 10 minutes, but maybe a weekend to allow for training.

E.g [1] trained for 20 * 10^5 iterations at 2.1 seconds per image, coming out to 48 days (!) in total. Maybe the 2.1 seconds also include the post-processing to get bounding boxes, though. Anyway, you need lots of training time.

[1] https://arxiv.org/abs/1604.04018 "Multi-Oriented Text Detection with Fully Convolutional Networks"

Looks like they only used a single GPU for training though - you could trivially get that time down a lot if you used one of the cloud ML platforms (and were prepared to pay)

Amazon Rekognition can almost do it. It doesn't seem to detect text, specifically, but it'd be able to detect whether or not a bird is in the image, or similar.

That's not what the project is asking for, but it's very close to being a very doable hobby project.

The problem is what birds look like is very well defined compared to text. As in, there are many types of birds, but each type has a "consistent image" but there are infinite variations of text

Fair enough. You could certainly still do it with Mechanical Turk then, although the delay in response might not meet the requester's unstated requirements.

Relevant XKCD: https://xkcd.com/1425

Reply on news.ycombinator.com