Hacker News Reader

Comment by raister | Hacker News Reader

How this gets to the front page and other relevant things get diluted amazes me every time in HN!

Lighten up. Have some fun.

Life is more than just "important" things.

Or better yet: "Important" and "relevant" things should not be limited to serious, constructive, work/society relevant things. Fun is very very important for the mind and soul.

Have more fun.

Today's friday. The site is cool and hilarious. Problem also looks like a good application for ML.

Bonus points for candid photos, no ads, low resolution images and nostalgic scenes.

I like this structure of HN. Kinda like old Slashdot. Something funny and lighthearted amongst all the heavy articles.

> Problem also looks like a good application for ML.

Wouldn't that be complete overkill, though? Zooming and panning are so much simpler and less compute intensive.

I think the idea was to find images that point to a certain location from a collection of photos.

Not having a bunch of photos of people pointing and just zooming/panning to that bit.

ML methods would require labelled data to be trained.

Labelling itself already solves the issue (i.e. a bunch of photos with people pointing in certain directions).

So basically you'd solve the problem you want to solve in order to automate a task that you already performed.

> ML methods would require labelled data to be trained.

Not necessarily.

One could train a model on synthesized data using e.g. blender, some programmable 3D people models with hand controls and some generic background images to paste them on.

I've heard of this being used. It's surprising to hear you mention it. Have you done that?

It's a neat idea, and I hope it becomes more popular.

> It's surprising to hear you mention it.

Me specifically or in general?

Anyway, for a Multimedia Information Retrieval course I chose to do my term project on training a neural network with synthetic data. In particular I modded Minecraft such that when I press a button it saves two screenshots: one regular and one where the game renders a depth map instead. I used this to generate ~1000 samples with perfectly accurate depth maps. Because of the mods, texture pack and world I used the data was somewhat realistic: https://i.stack.imgur.com/Zai51.jpg https://i.stack.imgur.com/eamMR.png

This data was then used to train a neural network to predict the depth map of unseen images. It was relatively successful, but requires more data and more research, I only had so much time for a term paper.

I guess people upvoted this submission because they think the website has some sort of machine learning algorithm, which is the daily bread on Hacker News.

Personally I don't think there's any machine learning needed for this (but who knows what they used). If I were to build a replica, my approach would be to grab a decent amount of suitable photos, manually mark the pointed areas in them, and have some sort of scaling/cropping/offset handle near misses.

I upvoted the submission because I think it's a cool trick either way.

It just shows that a "Lot of Prep" can be "machine learning" to the observer.

In the ye olden demoscene days, we called that "precalc".

Just calculate all the effects up front and then draw them "in real time" :)

If it were "lot of prep" - e.g. a manually-tagged and fixed database of about 100,000 images of people pointing (assuming 100 columns and 100 rows in the browser-viewport) then that's still small enough for an instantaneous response from the web-server, so the relative long busy-time suggests that _something_ computationally expensive is running in the background - such as a "this-pointing-finger-does-not-exist"-engine.

It could just be fake busy. Timing is everything in comedy, so give a first time user a little suspense to think advanced computer magic is happening before the punchline arrives.

Exactly. Looking at the JS code (even minified), you can see that it's calling a setTimeout for effect - possibly also waiting for image load-, but all the data is there already.

If it runs instantaneously, it would probably be a pretty cool effect to have the pointing fingers trace your cursor as it moves.

If I had a piece of ML that could do that, I'd just run it once on the server and grab a coffee as it generates a set of 100k labeled images for me.

IIRC, the "this X does not exist" sites also don't run ML on request, but serve cached images. At least that was my impression from getting the same image more than once when I spammed F5 fast enough.

All work and no fun makes hacker a dull person.

Reply on news.ycombinator.com