"I really see this as a start of the artificial intelligence boom in marine science" - first words of the video in the article.
I don't understand why to even say something like that. PR value? As I understood the paper, the presented method is very refreshingly clean of anything resembling what we call AI today. It seems to be a combination of good, old-school photogrammetry and image processing techniques - which is great, because with such methods they can actually ensure the result is physically correct.
In the paper [0] there is more info than just what's explained in the simplified blog post. A.I. is needed if you want to do the same without the color chart used for training it. Furthermore, it is not just color correction, also backscatter is taken into account.
Yes, I'm talking about this paper exactly. All I can see is photogrammetry and regular image processing stuff. There's no DNNs or other hot stuff. I guess the part where they're estimating parameters for the equations from photos counts as ML, but people have stopped calling this stuff "AI" quite a while ago.
I think it is more that this correction allows more generic algorithms like object classification to work.
Of course you could do object classification without color, but that would be throwing away information.
It is just that the back scattering currently is overwhelming the colors in such a variable way that it is hard to use color inclusive object classification without doing stuff like requiring the exact same distance between camera and object for each picture. And likely that doesn't even work yet.
> I guess the part where they're estimating parameters for the equations from photos counts as ML, but people have stopped calling this stuff "AI" quite a while ago.
Indeed- as that coined term suggests, "AI" has such a philosophically-tied definition with moving goalposts, that in practice it gets slapped far too liberally on anything 'magical' having to do with computation. (As long as it's still trendy to do so, at least.)
'Machine Learning' is at least a little more specific of a term (if still rather general).
People point at this all the time, but there's also another AI effect. People use traditional techniques to solve modern problems and label it "AI." Calling it a traditional technique despite the modern application doesn't mean it's moving the goalposts.
Aside from the paper, I think one thing they're referring to is using the output (color corrected images) as input for AI. One thing they want to do is count the number of fish in an image, and know which species each fish is. So you can take pictures of coral reefs and estimate "there's 1,000 species X, 2,200 species Y". With the old images, it's too difficult to determine which species a fish is. With the new images, it's easier. So Sea-Thru is preprocessing that'll be useful for AI in marine science.
I dont have the impression it works on moving objects, since it needs multiple frames from different depths, but it could be used to count static critters... unless some kind of boom with multiple cameras at different depths is used...
also, since the technique is removing a foggy haze, it seems like this could be used for selfdriving cars, with multiple cameras along the periphery of the car, to clean the image for foggy conditions (fog, smoke, smog, ...)
I think the idea is that if you can produce cleaner, physically-correct ocean imagery, then the result of that is more amenable to consumption by machine learning image processing tools.
It's sort of like how having a big image corpus enabled AI. There's no AI in the images themselves, but it gives you something to throw AI at.
I don't understand why to even say something like that. PR value? As I understood the paper, the presented method is very refreshingly clean of anything resembling what we call AI today. It seems to be a combination of good, old-school photogrammetry and image processing techniques - which is great, because with such methods they can actually ensure the result is physically correct.