[其它] Non-Local Image Reconstruction for Efficient BTF Synthesis

[复制链接]

彬彬

797 主题	1 听众	1万积分

资深设计师

Rank: 7 Rank: 7 Rank: 7

纳金币: 5568
精华: 0

电梯直达

楼主

发表于 2012-1-5 08:44:54 |只看该作者 |倒序浏览

1 Introduction

Virtual reproduction of optical material properties is an important

task for many computer graphics applications. Recently, bidirectional

texture functions (BTFs) [Dana et al. 1999] have become

quite popular for photo-realistic rendering of surfaces. BTFs can

represent a huge variety of different materials. They can be easily

integrated into any modern rendering system which offers programmable

surface shaders. BTFs of real materials can be optically

measured, they are known to be efficiently compressible and can be

used for real-time graphics rendering applications even on the web

using WebGL.

Alternatively, BTFs can be synthesized based on known microstructure.

This approach is especially interesting for computer

aided design of complex materials that cannot be represented by

simple BRDF or texture models. One challenging example is cloth

which exhibits complex single and multiple scattering of light between

millions of fibers. Cloth designers might want to edit the

optical properties of individual fibers, the geometric properties of

spun yarns and their overall composition using weave patterns or

knitting techniques and perform predictive rendering.

To this end, virtual gonio-reflectometer setups can be build which

compute the angular reflectance data for a virtual material sample.

Essentially, this means that light transport has to be simulated for

thousands of view and light directions. A naive solution would be

to build a virtual half-dome multi-view multi-light measurement

device and render e.g. 70 70 images (”BTF slices”) of a flat

material sample for 70 viewing and 70 lighting directions. Unfor-

tunately, rendering thousands of images, commonly done using a

path tracing approach, is extremely costly. To accelerate this process,

we take advantage of the fact that BTFs tend to contain many

very similar apparent bidirectional reflectance distribution functions

(ABRDFs). Exploiting this property, we reduce variance by

sharing radiance samples across different BTF texels.

In the context of image denoising, non-local means filtering techniques

[Buades et al. 2005] have become popular. Here, the original

value of a pixel is replaced by a weighted average of all pixels

in the image showing a similar distribution of surrounding pixels.

However, the technique breaks down for renderings with a very low

signal to noise ratio.