They create occluder meshes to partition a large NeRF into multiple smaller ones. That way, stepwise traversing through empty space in the NeRF can be avoided. But I'd guess that DepthNeRF would still be more efficient overall, because their partitioning might actually increase overdraw unless they reach an opaque alpha and the UE occlusion engine works pixel perfect.
Unbelievable!
I use WebODM for a few projects and was wondering when the team might consider incorporating NeRFs in their porduct. Seeing this done in UE4 and maybe some work i see great potential for construction work, farms or archeological digs in problematic areas for fast and only basic information-gathering needs.