Hacker News new | ask | show | jobs
by mensetmanusman 1695 days ago
AR camera shots often have an uncanny valley appearance.

It would be cool if there were an option to have the initial simulated image ‘annealed’ with a generative adversarial network.

Maybe you could also partner with local frame builders (ask for 2% of sales or something) to send clients your way, since, so often, the frame has a large effect on the art’s aesthetic impact.

2 comments

I think the uncanny valley thing is because the superimposition does not take into account the lighting conditions of the room.
Yup the superimposition is currently setup to ignore light in the room. I experimented with having it respond to light in the room, but ran into issues with the image rendering way too dark. I'll spend some more time fine tuning it and seeing if I can get more realistic light rendering into a future release
Isn't determining lighting without some reference card a "hard" problem? How do you tell the difference between a yellowish room illuminated by white (sunlight) versus a whitish room illuminated by warmer yellower bulbs?

I am a not a photography person so this a genuine question asked from ignorance

Yup that's a great question, I think generally speaking that is a really difficult problem in the AR space, which I think is part of the reason why most AR models still look fake or have the uncanny valley.

Using the ARKit libraries, it gets handled automatically for the most part. I think there's a lot of computer vision/image processing algorithms running in the background that potentially pick up on shadows/reflections in real world objects, which help it determine the white balance/color temperature of the room. I'm no expert on it myself though and not 100% sure on the science behind it

Go all the way and make a "buy now" option. That would entail partnering with printers and framers to offer standalone posters and framed versions of standard / custom artwork.