[yeldarb]

I browsed through the hundred or so other Startup School companies in my vertical and there’s nothing even close to my vision of the future (using computer-vision to add a software layer to real world objects; proof of concept: https://twitter.com/braddwyer/status/910030265006923776).

It’s pretty frustrating actually; it seems obvious to me that it will be the future of human-computer interaction.

I have been waiting for someone to build it for the past 18 months... but nobody seems to be working on it. I finally said “screw it” and I’m going to build it out myself.

If this sounds interesting feel free to reach out! I’m one of the ones looking for a cofounder. (Twitter is easiest. My dms are open)

[machinelearning]

Isn’t that the goal of every AR startup/project? Care to explain?

[yeldarb]

Maybe I haven’t been looking in the right places but I’ve downloaded most of the AR apps and games that have come out for iOS thus far and they’re almost all gimmicky.

Simply plopping a 3D model into the room isn’t particularly compelling or useful. But that seems to be the bar for “AR apps” today.

[machinelearning]

Most AR apps today are based on the same toolboxes, either opencv, arkit or arcore. The latter 2 are actually pretty good (in terms of tracking and plane detection) for more simple use cases.

So by gimmicky I assume you’re talking specifically about the use case?

If so it’s probably true that majority of ar apps are games that don’t work too well or ar just demos/prototypes.

However there are a few that are pretty functional e.g. the measuring app in iOS. Was your idea to consolidate all such use cases to create a general purpose app?

[yeldarb]

The magic comes when combining AR with an app that understands what it’s looking at and how to contextually interact with it.

The AR toolkits tell you about the 3D environment but they don’t do anything to tell you about the context.

When your app understands what it’s looking at it can add relevant functionality (eg solve a paper Sudoku puzzle, give hints, check your work, scan into a playable app, etc).

Happy to talk about it more! Shoot me a DM.

[machinelearning]

Generalized scene understanding is still a hard problem in both academia and industry. The big companies are working on such things with access to tons of data.

Assuming you do know what’s in the scene semantically most ar applications that aren’t games do intend to use that information. I’d say the primary hurdle there is still general scene understanding.

It’s easy to identify a sudoku puzzle but not so easy to identify and classify all 3D objects with any sense of precision yet. Seems like a large scale data play to me..

[yeldarb]

Definitely. Not everything I’d like to do is possible at the moment but Machine Learning is getting better so fast that more and more will become feasible bit by bit.

I have a bunch of ideas that seem like low hanging fruit, a bunch that seem hard but possible, and a lot that will be possible someday. And hopefully I can help accelerate that timeline.

[Vomzor]

You’re idea is to build that context layer? It seems the apps you’re talking about will become reality when we have useful AR glasses. But they will be manually built for each usecase.

[yeldarb]

Essentially. I have a list of use cases I think are possible to tackle now so I’m going to work on building out more single-niche apps in the near term.

In doing so I’ll have to build out tools and common code which I hope to eventually release as a platform to help other devs expand the idea out into other verticals.

[svantana]

Solving sudokus in AR could also be described as "gimmicky"...

[yeldarb]

Certainly. Proof of concept :)