This is incredible! I love how they went ahead and did work on not just compressing it, and also making it such that it is easily renderable on commodity platforms (basically making explicit geo and atlased textures, kind of like a regular game render)
Next step could be recognizing and generalizing shapes (human, car, house), deformation/movement models (rigging, inverse kinematics) and decompositing textures into layers (skin, clothes). You would keep common shape/movement/texture atlas in the base codec. As you wrote, the process of recording as a reverse game engine.