is there any way to inference in a long time video

When doing inference in a video, you need to use
`inference_state = predictor.init_state()`
to initialize the state, which will load all frames of the video, and if the video is large, it will take up a lot of memory or even fail.
I don't know why it needs to load all frames, is there any way to use segmentation or other methods to do inference in a large video?

Start by reading the predictor.init_state() entry point and reproduce its memory use with a large video. Trace why all frames are loaded, then determine what supported large-video inference should look like and how completion would be verified without exhausting memory.

Written by the indexing model from the issue text.