复杂场景真实感绘制中的高效采样方法研究

发布时间：2018-07-01 11:16

本文选题：真实感绘制 + 高效采样　；参考：《浙江大学》2016年博士论文

【摘要】：真实感绘制是计算机图形学中最重要的研究领域之一,它准确地模拟虚拟场景中光能传输过程,以呈现具有真实感的虚拟世界,并被广泛用于制造、设计和娱乐等行业。目前主要的真实感绘制方法都是基于积分方程的数值求解,即在光路的空间中随机生成采样并用采样加权求和方式得到积分方程近似解。由于光路空间维度很高并且非常复杂,因此这些方法需要巨大的计算量。本文针对复杂场景条件下的真实感绘制问题,以高效的绘制性能为目标,在提高采样效率方面沿着下述思路开展了研究工作：一个思路是通过提高采样的利用效率来提升绘制效率。在这个方向上,本文提出了两种基于光子映射的真实感绘制方法。第一种方法是针对全局光照明绘制,用辐射照度回归的方式提高最终收集中采样的利用效率。非参数的回归模型将光子映射和最终收集两种方法的优势相结合,并用光子映射的辐射照度预测最终收集的辐射照度,从而以接近光子映射的绘制代价得到接近最终收集的绘制质量。为解决预测失败的情况,本文提出了一种快速的异常样本剔除策略,并用最终收集算法计算这些异常样本的颜色。在相似的视觉效果下,本文算法相比最终收集算法获得了大约3倍的性能提升。在第二种方法中,针对非匀质半透明材质的高效绘制,通过在概率上调整光子采样的空间分布从而提高其利用效率。该方法基于体积光子映射算法,通过概率调整光子分布的方式有效地增加视域内的光子数,从而改进半透明材质绘制质量。针对概率调整引起的光子能量的不均衡问题,本文通过在距离上和方向上对光子进行切分,将高能量的光子细分成若干低能量的光子。在相同的绘制时间内,本文算法的绘制结果好于体积路径跟踪和体积光子映射的结果。另一个思路是提高单个采样的计算效率,本文将新的并行计算模式与GPU上的大规模并行计算相结合从而提升绘制效率。在这个方向上,本文提出了一种高效的具有可扩展性的可编程运动特效绘制算法。运动特效通过各种运动描绘风格有效地呈现物体运动信息,是真实感绘制中运动模糊的扩展。本文提出了一种GPU上的可编程运动特效并行绘制系统,采用时间和图像空间上的三维并行计算方式,大幅提高了线程间数据重用率,解决了GPU上可编程运动特效绘制时面临的并行度和显存容量的矛盾。为了解决轨迹线段之间的可见性冲突,本文还提出了一种基于区间切分的可见性算法。考虑到用户设计可编程运动特效的便利性,采用预计算像素上轨迹线段的方式,以进一步减少运动特效绘制的时间。
[Abstract]:Realistic rendering is one of the most important research fields in computer graphics. It accurately simulates the process of light energy transmission in virtual scene to present the virtual world with realistic sense, and is widely used in manufacturing, design and entertainment industries. At present, the main methods of realistic rendering are based on the numerical solution of integral equation, that is, random sampling is generated in the space of optical path and the approximate solution of integral equation is obtained by sampling weighted summation. Due to the high dimension and complexity of optical path space, these methods require huge computational complexity. Aiming at the problem of realistic rendering in complex scenes, this paper aims at efficient rendering performance. In the aspect of improving sampling efficiency, the following research work has been carried out: one idea is to improve the efficiency of rendering by improving the efficiency of sampling. In this direction, we propose two realistic rendering methods based on photon mapping. The first method is to improve the efficiency of sampling in the final collection by using the method of radiance regression for global illumination rendering. The non-parametric regression model combines the advantages of photon mapping and final collection, and uses the radiance of photon mapping to predict the final collected illuminance. Thus the rendering quality of the final collection can be obtained at the cost of close to photon mapping. In order to solve the problem of prediction failure, this paper proposes a fast exception sample elimination strategy, and calculates the color of these abnormal samples by the final collection algorithm. Under similar visual effects, the performance of the proposed algorithm is about three times higher than that of the final collection algorithm. In the second method, aiming at the efficient rendering of non-homogeneous translucent materials, the spatial distribution of photon sampling is adjusted in probability to improve its utilization efficiency. Based on the volume photon mapping algorithm, this method can effectively increase the number of photons in the field of view by adjusting the distribution of photons in probability, thus improving the rendering quality of translucent materials. In order to solve the problem of photon energy imbalance caused by probability adjustment, the high energy photons are subdivided into some low energy photons by dividing them in distance and direction. In the same rendering time, the proposed algorithm is better than the volume path tracking and volume photon mapping results. Another idea is to improve the computational efficiency of a single sample. This paper combines the new parallel computing model with the large-scale parallel computing on GPU to improve the rendering efficiency. In this direction, this paper proposes an efficient and scalable programmable motion rendering algorithm. Motion effects represent object motion information effectively through various motion depiction styles, which is an extension of motion blur in realistic rendering. In this paper, a programmable motion special effect parallel rendering system on GPU is proposed. The method of 3D parallel computing in time and image space is used to greatly improve the data reuse rate between threads. The contradiction between parallelism and memory capacity in programmable motion effect rendering on GPU is solved. In order to solve the visibility conflict between trajectory segments, a visibility algorithm based on interval segmentation is proposed. Considering the convenience of the user in designing programmable motion effects, the trajectory segment on the pixel is predicted to further reduce the time of motion special effect drawing.
【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】