Fast De-occlusion Neural Radiation Field

Neural Radiance Fields (NeRF) has demonstrated remarkable capabilities in the field of novel view synthesis, but most studies are limited to static scenes. However, in practical applications, outdoor scenes with transient objects are more common. In these scenes, the scene to be reconstructed is often not isolated, but accompanied by various transient objects, such as moving objects or occluders. Given that most current NeRF models mainly learn static objects through perceptual networks, they face challenges in dealing with transient objects. This paper proposes an optimization model that combines a transient MLP perceptron network specifically for learning transient objects. It also introduces a hybrid effective sampling and key sampling method and a cache structure based on a double-layer linear table to accelerate model rendering speed. The density information output by the transient MLP network can change with the change of the observation angle, and an additional uncertainty is output, thereby achieving effective learning and screening of transient objects. This helps reduce the impact of occluders on the reconstruction results.