We propose to learn NSLF alongside the real-time reconstruction. Different from NeRF, NSLF does not relies on ray-integration and thus the spatial training could be in parallel spatially. Then, we manipulate the NSLF by Multiple Asynchronous Neural Agents (MANA) to allocate agents and distribute data. And construct a unbounded and high efficient NSLF-OL model. It has potential to work in real-time on arbitrary large scenes.
Quick online demo please find our code. :D