Neural Rendering of Dynamic Scenes

In this neural rendering project, by adding the explicit module of scene flow fields to the original NeRF, a long time of dynamic 3D scenes can be represented.

Reproduce the Scene with Fixed Time.

Reproduce the Scene with Fixed Viewpoint.

The sudden appearance of the vehicle in the right image is due to the fact that the original camera view was also moving forward and the rear of the vehicle was not visible at some moments.

This method is more like the 3D version of video compression, where the canonical 3D scene at the original time corresponds to the reference frame in the compressed video, and the implicit or explicit scene flow fields correspond to the motion vector.