光线跟踪绘制中的光路重用技术研究

发布时间：2019-05-17 03:41

【摘要】：真实感图形绘制是计算机图形学最重要的问题之一,而蒙特卡洛光线跟踪是真实感图形绘制中最主流的方法。该方法可以处理任意的光源类型、场景表达、纹理材质、以及照相机设置。其缺点是计算复杂度较高,绘制性能较低。光路重用是提高蒙特卡洛光线跟踪算法性能的重要途径。通过重用一部分光路,可以将一条光路多次用于光能传输的估计,从而降低采样的平均成本。另一方面,通过选择较为重要的光路进行重用,也可以实现一定程度的重要性采样。根据重用光路采样类型的不同,光路重用技术主要分为三类。由视线或光线方向生成的光路出发,重用反向光路进行连接,从而生成一条完整光路的,称为双向光路重用。将原光路采样前半部分与另一条同向光路后半部分的光路进行连接生成完整光路的,称为单向光路重用。在二维图像空间内,对光路采样的估计结果,也就是着色值,进行直接重用的,称为图像空间着色重用,也等价于对于整条光路的直接重用。本文系统地研究了蒙特卡洛光线跟踪框架中的光路重用问题。三个主要工作涵盖了三种不同类型的光路重用,分别是解决了双向光路重用中的估计偏差问题,分析并减轻了单向光路重用中协方差对于估计方差的影响,以及利用二维图像空间中光路采样结果的一致性提出了高性能的毛发绘制算法。具体来说,本文包含如下内容:(1)针对双向光路重用,我们提出了一个新的光子采集方法来替代之前光子映射算法中的密度估计,从而得到无偏的绘制结果。我们会单独处理每一个光子,将光子所代表的光线与进行采集的视线相连,来形成一条完整的光路。这个连接可以被理解为一次俄罗斯轮盘事件,其成功概率是一个积分。由于这个连接的概率出现在最终贡献估计的分母上,人们一直没有高效的方法来避免在估计概率积分的同时引入偏差。因此我们提出了一个无偏的算法,利用一系列伯努利试样来对这个概率积分的倒数进行无偏估计。最终的方法是一个完备的无偏采样技术,并且可以被结合到广泛的蒙特卡洛光能传递算法之中。(2)针对单向光路重用,我们研究了该类方法中光路重用的协方差对于最终绘制方差的影响。通过我们的分析可以发现,在此类方法中估计方差由方差项以及协方差项两部分组成。我们发现提高重建采样率只能降低方差项部分,而对于协方差无能为力。我们也发现协方差项实际上表示了间接光照采样在重建过程中被重用的程度。同一个间接光照采样被越多的重建采样所使用,其带来的协方差项也就越显著。基于对于方差的分析,我们提出了一个自适应生成间接光照采样,从而降低协方差项的方法。我们将这个自适应的方法应用到间接光场重建算法以及轴对齐的图像空间滤波算法之中,并显著地减轻了绘制结果中的失真。(3)最后,针对图像空间的着色重用,我们提出一个基于光锥的光线跟踪技术,来进行毛发物体的高质量绘制。通过将一个像素内的所有光线聚合成一个光锥,我们实际上可以在这个光锥内对于每个采样的计算结果进行重用,以此来降低毛发的纤细几何所需要的巨量采样。这样一来不仅减少了采样的计算代价,也减少了半透明的毛发所需要的合成计算的代价。最后的结果是一个高效的光线跟踪算法,可以生成与其他方法同质量的绘制结果,同时显著减少计算的时间。
[Abstract]:Realistic image drawing is one of the most important problems in computer graphics, and Monte Carlo ray tracing is the most important method in the drawing of real-sense graphics. The method can process any type of light source, scene representation, texture material, and camera settings. The method has the advantages that the calculation complexity is high, and the drawing performance is low. The optical path reuse is an important way to improve the performance of the Monte Carlo ray tracing algorithm. By reusing a portion of the optical path, an optical path can be used multiple times for the estimation of the optical energy transmission, thereby reducing the average cost of the sample. On the other hand, by selecting a more important optical path for reuse, it is also possible to realize a certain degree of importance sampling. According to the different sampling types of the reused optical path, the optical path reuse technology is mainly divided into three types. And the optical path generated by the line-of-sight or the optical line direction is used for connecting the back light path so as to generate a complete optical path, which is called a two-way optical path re-use. And the optical path of the half part of the front half part of the original optical path is connected with the optical path of the half part of the other co-directional light path to generate a complete optical path, and is called a one-way optical path reuse. In the two-dimensional image space, the estimation result of the optical path sampling, that is, the coloring value, is directly reused, called image space color reuse, and is also equivalent to the direct re-use of the whole optical path. In this paper, the problem of optical path re-use in the Monte Carlo ray tracing framework is studied systematically. The three main work covers three different types of optical path re-use, respectively solving the estimation deviation problem in the two-way optical path re-use, analyzing and reducing the influence of the covariance on the estimation variance in the one-way optical path re-use, And a high-performance hair drawing algorithm is provided by utilizing the consistency of the light path sampling results in the two-dimensional image space. In particular, this paper includes the following contents: (1) For bi-directional optical path re-use, we propose a new photon collection method to replace the density estimation in the previous photon mapping algorithm, so as to obtain an unbiased drawing result. We will process each photon separately, and connect the light represented by the photon with the line of sight to be collected to form a complete optical path. This connection can be understood as a Russian roulette event whose success probability is an integral. Since the probability of this connection appears on the denominator of the final contribution estimate, there has been no efficient way to avoid introducing a deviation while estimating the probability integral. So we put forward an unbiased algorithm that uses a series of Bernoulli samples to unbiased estimate the reciprocal of this probability integral. The final approach is a complete unbiased sampling technique and can be incorporated into a wide range of Monte Carlo radiosity algorithms. (2) For one-way optical path re-use, we have studied the influence of the covariance of the optical path re-use on the final drawing variance in this kind of method. By our analysis it can be found that, in such a method, the estimated variance is comprised of a variance term and a covariance term. We have found that increasing the reconstruction sampling rate can only reduce the variance term portion, and there is nothing to do with the covariance. We also found that the covariance term actually represents the extent to which indirect illumination samples are reused in the reconstruction process. The more the same indirect illumination sample is used by the more reconstruction samples, the more significant the covariance term is. Based on the analysis of variance, we propose an adaptive generation of indirect illumination sampling to reduce the covariance term. We applied this adaptive method to the indirect light field reconstruction algorithm and the axis-aligned image spatial filtering algorithm, and significantly reduced the distortion in the rendering results. (3) Finally, aiming at the color reuse of the image space, we present a light-cone-based ray tracing technique for high-quality rendering of hair objects. By aggregating all the light in one pixel into a light cone, we can actually reuse the calculated results for each sample within this cone to reduce the large amount of sample required for the fine geometry of the hair. So that not only the calculation cost of the sampling is reduced, but also the cost of the synthesis calculation required by the semi-transparent hair is reduced. The final result is a high-efficiency ray tracing algorithm that can generate the same quality rendering results as other methods, while significantly reducing the calculated time.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】