Explicit Composition of Neural Radiance Fields by Learning an Occlusion Field

The neural radiance field (NeRF) is an implicit representation of the appearance and shape of objects based on neural networks. While numerous studies have shown the great performance of NeRF on tasks like free-view synthesis, it remains a challenge to composite multiple objects depicted by NeRF models. Different from previous works that focus on decomposing separate scenes from a compositional NeRF, we study how to naturally composite two trained NeRF models without combined scene images for supervision. Specifically, we propose a novel framework that learns an occlusion field (OCF) to model the occupancy property of the object. A dedicated compositional rendering equation and loss functions are then designed to learn an accurate occlusion field. With the trained occlusion field, the source object represented by NeRF can be explicitly scaled, moved, and merged into the target NeRF. Our method can synthesize plausible merged results with accurate occlusion relation and natural transition at occluding boundaries, and even work for challenging objects like hairs and leaf clusters. Experiments show that our method is superior to the previous methods, providing a new solution for the efficient and effective composition of NeRF models.