A few people commented that the foreground/background detection cannot keep up with movements fast enough. Here's an idea that might help, although I'm not sure

A few people commented that the foreground/background detection cannot keep up with movements fast enough. Here's an idea that might help, although I'm not sure if it can realistically be done: When the video is encoded, the codec does motion estimation (among other things) to reduce the bandwidth required. So why don't we use the motion vectors from the video codec to modify the foreground/background mask in real time? Obviously this is going to create weird artifacts pretty soon, but it might just be good enough for a few frames before the ML model produces another accurate mask.