Hacker News Reader

Comment by ivanech | Hacker News Reader

actual project site with more detailed description and before+after images: https://www.deryaakkaynak.com/sea-thru

paper [pdf]: http://openaccess.thecvf.com/content_CVPR_2019/papers/Akkayn...

"Sea-thru" is such a great name for this. Also I'm always glad to see things that aren't "we threw a neural net at it and it looks pretty good"

Soundtrack: https://youtu.be/NWB02uS2Gow

The company name ("SeaErra") is pretty amusing too :)

I just wish that site didn't require JavaScript just to view these images. How is this Wix behavior acceptable?

Common users don't care or don't even know about javascript.

that's fascinating, and a lot more involved than what I expected (my guess was something more like off-the-shelf color correction)

The palette cards they carry around remind me of white-balancing.

If I understand things correctly it's not entirely different to white balancing either, just that the difference between white balancing a monitor or camera is that you don't really have to correct for the scattering of light through air because it's negligible at the distances you're likely to work with. The trick here – again, if I understand correctly – is that this method also finds the light scattering/absorption characteristics and corrects for those as well. It might seem obvious, but I think it's more involving than it seems, because light scatters differently in different environments and I think – again, I might be wrong, this is just my understanding – this method accounts for that by using multiple images at different distances to gather the parameters to find the scattering and absorption characteristics. This is what makes it physically accurate and not just "make these colors mostly look good" which is really quite cool and useful!

If I remember correctly, the percentage of light that passes through a column of water exponentially decays with distance, like stacking a bunch of fixed-percentage optical filters on top of each other. They’re using parallax shift between multiple images to generate a depth map, so they have a measure of the amount of water obscuring each pixel.

I’m not sure how they’re finding the optical properties of the water; there’s something about using the darkest pixels in the image for calibration that I didn’t quite follow.

It's actually exactly that. Her set of skills is to do that actually accurately with setpoints for both distance to the objects as well for different wavelengths simulataneously.

On the first example corrected image reveals some details in shadows but also burns out some highlights.

So it seems like it's just MVS > depth map + recovery of haze characteristics by sampling a color reference at a bunch of distances. That seems blindingly obvious unless I missed something.

Don't tell me, you could have done it yourself in your basement in a couple of hours? Well done.

This is the most Hacker News comment I've seen in a really long time

Good job, dude

It's so HN it's almost /.?

You missed something: the huge amounts of effort required to make it work and produce acceptable results

And that no one had done it before (apparently).

First of all, you're wrong. "It works without a color chart, or any information about the optical qualities or depth of the water body."

Secondly, if you think it's "blindingly obvious", read the paper linked in the parent's comment. Even a quick skim will show you're wrong.

In conclusion: yes, you missed something.

Reply on news.ycombinator.com