The term rendering refers to a set of techniques used to generate 2D projections from 3D data of a scene. Most multi view image synthesis pipelines can be divided into 2 blocks. First, scene representation block which constructs a 3D scene representation from multi view images of a given scene. Second, a volume rendering block which is used to generate novel view images of the scene from the learnt 3D scene representation. Depending on the type of representation used in the first block, there exist several volume rendering approaches. In this blog, we explain a radiance field based volume rendering approach.

Rendering multiview images of a scene has been a long standing research area in the field of computer vision. Given multiple images of a scene with varying viewing directions, the task of multiview image synthesis is to generate photorealistic images of the scene with novel views.