Time Prediction For Reyes Rendering Architecture Based On Adaboost. MH Algorithm

With the development of multimedia technology and CG technology, rendering engine has become more and more widely used in film animation, simulation and game effects. At the same time, with the increase of rendering computation, it became difficult to meet the huge amounts of computation with a single computer. As a result, cluster rendering came into being. Cluster rendering platform mainly adopts RenderMan, mental ray and Vray rendering engine to provide rendering services, these rendering software using Reyes (Renders Everything You Ever Saw) rendering algorithm, which can realize fast high quality rendering.Although Reyes rendering speed is faster than other algorithms’, but in a render farm, one of the key issues is the strategy of scheduling and dispatching rendering jobs, which greatly affects the computing efficiency. Time prediction for a render job plays an important and essential role in the job scheduling and dispatching stage. In addition, rendering cost is based on time prediction. However, there is no feasible algorithm and even little research work on this problem.Based on the requirement for efficient parallel of cluster rendering, this paper focus on Reyes rendering architecture according to RenderMan interface standards. We first analyze the factors that affect the rendering time and extract Pixel Sample, pixel number, light number, shadow number, material number, subdivide patch number and thread number seven key features as the feature vector based on the analysis. Then we propose a time prediction framework based on AdaBoost.MH algorithm, in which we transform the rendering time into intervals and combine them with the feature vector to obtain the samples. Furthermore, the experimental results show the effectiveness of the algorithm, and the accuracy of training set and test set is79%and78%.This paper shows the main contribution:1. Analyzed Reyes rendering architecture in detail and found the factors that affect the rendering time, through experiments to extract seven key features as the feature vector, then ready for the use of machine learning methods.2. Although time estimates is regression problems from general sense, the output should be continuous value, but the estimate of time does not necessarily require very accurate, as long as within a certain error, you can accept it. This paper transformed time forecast problem to classification problem, and applied to the AdaBoost.MH algorithm, then got the time estimates range, thus found a new application for AdaBoost.MH algorithm in rendering time predition.3. Proposed a rendering time predition framework in Reyes rendering architecture, which laid the foundation for the parallel of cluster rendering and rendering cost.This paper needs more works on the feature extraction of material types in the future.