[2021]GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields

NeRF in , We use a radiation field to simulate a real scene , Although the camera can change its position at will , But the things in the scene are fixed , Because a radiation field is a whole . In this paper, multiple radiation fields are combined , Thus, any radiation field can move freely in the scene . Of course , In this paper, the radiation field radiance field be called feature field.

The composition of the article ：

1. First of all, the whole process is used GRAF That set , So it includes a generator and a discriminator , But there are differences inside the generator . One side GIRAFFE There are multiple radiation fields in , On the other hand, a little change has been made in rendering .

2. First , Because there are multiple radiation fields , Therefore, each radiation field has its own set of Networks , Corresponding to different positions , Shape and texture parameters .

3. The second is how to combine multiple radiation fields , It's easy , Is along the light , Sum the colors according to the density of each object .

4. Finally, how to render ,GRAF It's done directly with volume rendering , But this paper adds a neural rendering after volume rendering , It is estimated that it is to improve the image quality , It is equivalent to refining the picture . It is worth mentioning that , The result of volume rendering in this paper is not a graph in the usual sense , It's a feature map , Because it does not output color , It is a characteristic value . Personally think that , The reason for this is that there is also a neural renderer behind , Therefore, some of the expression ability is reserved .

5. The last discriminator and GRAF The same as in China , It is a network that simply judges whether the output image is true or synthetic .