AI slow motion by nVidia

[SandalsWarrior]

Very cool project from nVidia, they used AI to interpolate between frames to create "fake" slow motion, it looks pretty convincing though. 

 

[3 Lions]

Unfortunately this technique is only decent when the video was filmed on a high speed camera in the first place.When you take a 60 FPS clip and put this effect on it, it just looks off.  

Edited June 19, 2018 by LinusTechTipsFanFromDarlo

[xAcid9]

Do you actually need AI to be able to do this?

[SandalsWarrior]

They are taking a 30 or 60 fps video, not a high frame rate one. My guess is that it has to be a decent resolution to have enough tracking points to interpolate from

[SandalsWarrior]

1 minute ago, xAcid9 said:

Do you actually need AI to be able to do this?

yes, to achieve this effect you need a high speed camera, they used a basic 30 fps video. It's like taking a video that you filmed with your phone  at 30 fps and getting extra frames out of it

[Nicnac]

siccc

 

kinda scary to think about watching a video where there are let's say only 1 real frame for every 3 cgi frames... that's a 3/4 generated video and not real but still real..... yeah i'm confused haha

[SandalsWarrior]

2 minutes ago, Nicnac said:

siccc

thicc

[Drak3]

10 minutes ago, xAcid9 said:

Do you actually need AI to be able to do this?

It's not actually AI. Nearly nothing claimed to be AI is.

[SandalsWarrior]

1 minute ago, VegetableStu said:

lol, step through the video frame by frame (using < and >). dat painter's melt ._.

i rate this similarly to fake bokeh

it does look a bit off in some examples but I wouldn't be surprised to sell this technology to other companies(i.e. a phone company) to use it in a flagship phone or smth(after Samsung did the 960 fps thingy, a few companies took interest in slow motion(mostly to be competitive) and I can see them use this like Google uses AI to process images in Pixel phones)

[Jurrunio]

4 minutes ago, VegetableStu said:

yeah actually I'd like to see anime scenes being processed come to think of it ?

yup, those 29.97fps slideshows hurt the experience

[SandalsWarrior]

8 minutes ago, Drak3 said:

It's not actually AI. Nearly nothing claimed to be AI is.

ofc it's not, because we don't have true AI yet, but no one calls narrow AI "narrow AI", it's clear that we are talking about narrow AI

[SandalsWarrior]

8 minutes ago, VegetableStu said:

yeah actually I'd like to see anime scenes being processed come to think of it ?

there's so much you can fill in the blanks with. it'll never be the same as dedicated hardware, although one could argue the plebs can have a pleb version functionality with a pleb camera

 

I hope in the next 10 years we don't forget what reality should look like ._.

agree, it won't replace a high speed camera but I see it exactly like the AI used in Pixel phones, to enhance lower end hardware and get more out of it through software. Totally fine by me

[3 Lions]

4 minutes ago, SandalsWarrior said:

Samsung did the 960 fps thingy

The Sony Xperia XZ Premium was actually the first phone on the market to do something like this. The thing is though that it's not AI they actually have miniature high speed cameras built into the phone, they have their own super fast memory which is I think closely related to HBM. 

[Drak3]

2 minutes ago, SandalsWarrior said:

ofc it's not, because we don't have true AI yet, but no one calls narrow AI "narrow AI", it's clear that we are talking about narrow AI

We're not talking about AI at all. It's a (complex) set of algorithms that do input -> output.

[Jurrunio]

2 minutes ago, VegetableStu said:

30?! more like 10 to 20 in most cases!

oh really? my TV source do send 30fps to my TV. Looks worse than video captured videos anyway

[SandalsWarrior]

2 minutes ago, Drak3 said:

We're not talking about AI at all. It's a (complex) set of algorithms that do input -> output.

I am not extremely familiar with the more advanced types of AI, but I am pretty sure that they used a trained neural network to do image processing. There are no algorithms to do image interpolation as convincing as this example

[BuckGup]

Just looks like they used optical flow and I wouldn't call this AI. Machine learning or just an algorithm at best. We have had stuff like this for years

[Deus Voltage]

11 minutes ago, Drak3 said:

We're not talking about AI at all. It's a (complex) set of algorithms that do input -> output.

Incorrect. There are (very roughly speaking) three kinds of AI:

 

-Narrow AI (as the OP pointed out)

-General AI (we're still, according to estimates by the likes of Ray Kurzweil, 10 to 20 years away from it)

-Super AI (the debate on how long it would take for super AI to become a reality is ongoing, but there is an agreement that it would ramp up rapidly after achieving General AI).

 

What you have in your calculator, phone, ect... is "narrow AI". The problem seems to have a linguistic dimension, as the word "AI" is used so often that the currency of the exchange is somewhat diluted. However, for the time being, stating "narrow AI" is sufficient for algorithms that are specialized in one domain. How those algorithms were achieved is irrelevant (case in point, Alphago Zero vs Alphago Master)

[Drak3]

7 minutes ago, SandalsWarrior said:

am pretty sure that they used a trained neural network to do image processing.

That's just a specialized processor/network of processors.

7 minutes ago, SandalsWarrior said:

There are no algorithms to do image interpolation as convincing as this example

On the market.

[Drak3]

7 minutes ago, Deus Voltage said:

Incorrect. There are (very roughly speaking) three kinds of AI:

 

-Narrow AI (as the OP pointed out)

-General AI (we're still, according to estimates by the likes of Ray Kurzweil, 10 to 20 years away from it)

-Super AI (the debate on how long it would take for super AI to become a reality is ongoing, but there is an agreement that it would ramp up rapidly after achieving General AI).

 

What you have in your calculator, phone, ect... is "narrow AI". The problem seems to have a linguistic dimension, as the word "AI" is used so often that the currency of the exchange is somewhat diluted. However, for the time being, stating "narrow AI" is sufficient for algorithms that are specialized in one domain. How those algorithms were achieved is irrelevant (case in point, Alphago Zero vs Alphago Master)

That's a load of marketting bullshit.

 

It's an algorithm set with fixed function. Nothing more.

[LAwLz]

6 minutes ago, Drak3 said:

That's a load of marketting bullshit.

 

It's an algorithm set with fixed function. Nothing more.

If it's a neutral network, which I think this is, then it's not a "fixed function algorithm".

 

 

 

 

Also, who the hell cares if it fits your (not strictly referencing you Drak3) definition of AI or not? Can we please stop arguing word choice and instead talk about this technology? It's pretty cool, but some people seem more occupied with being pedantic and nitpicking words.

This reminds me of waifu2x, which uses a convolutional neural network for upscaling and denoising pictures (specifically trained for anime characters).

[Deus Voltage]

13 minutes ago, Drak3 said:

That's a load of marketting bullshit.

 

It's an algorithm set with fixed function. Nothing more.

Incorrect/ I sense an emotional response. I advise pragmatism and I invite you to commit to doing academic/scholarly research (as all in this forum should). 

 

AI is not marketing. On the other hand, The terminator/ Ex Machina/ Transcendence/ 2001: a Space Odyssey/ ect... are in some vague sense "marketing". 

 

I advise you to watch scholarly and parliamentary debates in which Nick Bostrom and other other AI experts such as Google's Ray Kurzweil are involved. Hawking's views on AI are worth paying attention to for the sake of balance and nuance.  

 

Sources:

 

Nick Bostrom on AI in UK parlimant:

https://www.youtube.com/watch?v=wDU-fnXIV0s

 

US AI hearings:

https://www.youtube.com/watch?v=DZ4058m0a_g

 

Stephen Hawking on AI:

https://www.youtube.com/watch?v=fFLVyWBDTfo

 

[Froody129]

This is actually really cool. I wonder how the processing gets done though, and what kind of hardware you would need. Is this like a Titan V usecase or something you can do on a phone?

[xAcid9]

I mean we have re-projection in VR which is basically a real time frame interpolation iinm? This look similar but at 4x and higher frames.

[SandalsWarrior]

1 hour ago, Froody129 said:

This is actually really cool. I wonder how the processing gets done though, and what kind of hardware you would need. Is this like a Titan V usecase or something you can do on a phone?

I would guess they used something like a Titan V for training but after the network was trained, I don't think they need that much processing power for actually processing the frames