A robot or vehicle building a 3D model of its surroundings as it moves has to send a continuous image stream to a server that does the heavy computation. Unlike offline reconstruction, where you can gather everything first and think later, camera poses and geometry are estimated while the platform is still moving, so multi-view consistency becomes a real-time requirement rather than a post-processing concern.

That makes the network link part of the perception system. A dropped or delayed frame is not just missing data, it distorts the geometry being estimated right now, and the authors note how sensitive that estimation is to communication-induced problems.

Semantic communication is the proposed response: transmit what the reconstruction needs rather than the pixels a camera happened to capture. If the server is going to extract geometric structure anyway, sending a faithful image and then discarding most of it is a strange use of scarce bandwidth.