|
发表于 2018-11-14 10:38:04
|
显示全部楼层
Feature statistics: We compute the mean and standard deviationfor the K = 5 additional features of all samples in the pixel. Tocapture more global statistics, we also calculate the mean and standarddeviation of the pixel-averaged features in a 7×7 block aroundeach pixel. We compute the statistics for each component (e.g., i, j,k for shading normal) separately and average them together to geta single value per feature. Thus, we have 20 total values for eachpixel and the block around it.
The input layer consists of 36 nodes corresponding to our secondary features, while our output layer has one node for each filter parameter (6 total).
你的疑问来自这两段么?
第一个是关于输入数据的处理,具体内容文章已经有了。
第二个是Hidden Layer的层数,虽然相关但是是两个概念,HL可以更多但是失去了意义。比如对于Pure Diffusing场景,是没有任何Specular/Glossy的,这个时候照明复杂度的阶数降低,需要的Primary Feature更少,相应的Pass也更少,可能HL的数目也会更少。
这个Denoising技术有现实的买主,就是皮克斯的这个龟速渲染器。
https://renderman.pixar.com/reso ... _20/risDenoise.html
注意看Channel,在看论文中的Figure 4,那个就是对应的Primary Feature。讽刺的是用了这个之后这个破玩意还是慢的要死,出的图一片模糊。讽刺的是许多艺术家用后期合成的方法实现了Denoising,都不需要任何这些所谓的算法。
|
|