How exactly does importance sampling works?

[malfonsoarquimea]

I have 2 doubts regarding importance sampling.
The first one is about how much rays do you pick per iteration. If i understood it correctly, you would pick as many rays as there are pixels for a given frame (accounting for all the views), as a regular nerf would do, but sampling according to the generated weights instead of simply shuffling all the rays. So, to make it clear, if there are 10 views and each view is 100x100 pixels, a nerf would pick all 100x100x10 rays in a random order and DyNeRF instead would pick 100x100x10 rays (the same amount) with each ray having a probability of being picked proportional to its weight (computed using global median or temporal difference). Is this right?

The second one is about temporal difference. If we have 300 frames and we use a frame difference of 25, for a frame at time T we would measure the difference between that and the one at T+25 and compute the weight for frame T. But what about frames from 275 to 300?

Thanks in advance, any help you can provide would be greatly appreciated!