I guess this [0] might be a decent starting point. Also consider size comparisons of various gaussian splatting techniques (ex [1] & [2]). Static 3D scenes aren't technically the same thing as a 2D time series of frames, but since videos are generally composed of highly spatially related data there's bound to be overlap (ex [1] has already been extended to support dynamic scenes).
And keep in mind that security footage has an awful lot of both static and repetitive objects. It's an ideal target for more complicated approaches to compression.
And keep in mind that security footage has an awful lot of both static and repetitive objects. It's an ideal target for more complicated approaches to compression.
[0] https://github.com/ppingzhang/Awesome-Deep-Learning-Based-Vi...
[1] https://maincold2.github.io/c3dgs/
[2] https://jasonlsc.github.io/nerfcodec_homepage/