Maxine and Super Slow Motion?

AI & Data Science Deep Learning (Training & Inference) Maxine
1

Hello everyone,

will Maxine include Super Slow Motion functionality at some point?
Is this planned? Or will this functionality transition to another project (I came here from NGX)?

cu
Michi

2

Hi Michael,
We haven’t announced any super slow motion capability as a part of Maxine yet. We’re exploring many of the challenges around the quality of audio and video as well as efficiency of moving video and audio over networks. If this is something of interest, I’m happy to put in the request.

3

Hello Jeremy,

for me this topic is currently exploratory. The super slomo demo videos back then were very impressive and I also work with sports video so having some smoother slow motion could be nice. I have only access to a Quadro RTX 4000 with Linux, so I couldn’t use the NGX SDK as it was Windows-only. Now that NGX development has stopped and users are referred to Maxine and Maxine being available on Linux I wanted to give it a try, check the quality, stability, speed on the RTX4000, performance scaling with resolution, etc. But if it’s not available, that’s ok.
The other thing is that I heard one key prerequisite to a good AI solution is a good training set. So generating this training set for the sports discipline you are working with will also be an issue as you would need to invest into a high frame rate camera to shoot your training data - a cost that you may have wanted to avoid in the first place. ;)

But you were mentioning that Maxine focuses on challenges around quality of audio and video over network connections. As an amateur in the AI area it seems to me that in this context slow motion AI technologies may have a place? You shoot the webcam video with the normal 30 or 60 fps which will be your ground truth. Then you downsample the video to 15 or 10 fps and train the AI to interpolate the missing frames. Once done, you can reduce the frame rate of video connections over the network because you have an AI on the other side that restores the original frame rate. This could even be real-time trained as the sending side has access to both ground truth and reduced frame rate video and can fine-tune the network to the specific video source. But I guess if this is really a thing someone must have figured this out already. :)
And it should actually be part of the subject: AI-assisted video codec.

Thanks for coming back to me, I will hang around and see whether this slomo technology becomes available at some point.

cu
Michi