Video Compression

Sharon Carmel : CEO, Beamr
Slides
1B camera in the world, but this is boring to watch
from 2022 : 82 years of video uploaded on YouTube Every day
Now videos and images are structured data (this was no the case before
1Mh to train a model to drive a car (multiple resolution, different FPS)
Vidéo are a large data type, so heavy to compute
VPU : Video Processing Unit (Decode, encode video)
The encoder knows how much action is at which place in the frame
THe model will give hints to the encoder to have better compression
NVENC : some compromise compare to Software encoding because the silicon is limited
Quanlity & Performance 1080p : AV1 previous standard
The same NVENC can now improve quality with Driver update (3 years of work)
CABR : Content Adaptive Bit Rate
Since CUDA 12.1 and 12.2 (even on T4, L40, L40s, RTX 4000)
The model integrity in kept when training autonomous cars with CABR
Need to develop a lot of Guardrails for software enconding (for difficult frames)
The mathematical efficency of the current encoders is very high, but we can make it better