Boob friendly video compression layer

One method for lossy video compression is to take two frames that are duplicates or near duplicates and merge them, saving only one frame. For example if a person is talking in front of a wall, there's no need to save every single frame of that wall. You can take one picture of it and just use that until the lighting changes substantially or the camera moves. That's all well and good.

But for humans some parts of an image are much more important than others. A person's eyes may be in nearly the exact same position, but a couple of millimeters of movement there is important to communication. You can tell whether someone is looking at your eyes or at your nose from across the room.

Modern compression algorithms tend to throw out the baby with the bathwater. They'll happily compress the wall with the same vigor as the unmoving sword in a swordfight. A tiny twitch of the sword is important, but a slight shadow on the wall is not.

So the idea is to add to video compression a recognition of the most important parts of a video. All human or animal movements are important no matter how small. But set pieces are much less so. Face recognition is already done by cameras for the purpose of focus and so forth. This video compression library would add thousands of probably-important general shapes to mark for less lossy compression, allowing higher effective fidelity at lower file sizes.