Abstract:
The article presents an image preprocessing method for differentiable rendering tasks based on the analysis of the geometry of a scene and a series of images of an object from different camera angles. The proposed approach is aimed at reducing the amount of data transferred to differentiable rendering to calculate gradients based on scene parameters. In addition, the possibility of eliminating noise resulting from poor-quality measurements of the luminance of a real-world object or when creating synthetic images with a small number of rendering steps is being considered. This paper describes an algorithm for selecting the most informative points of a scene and an algorithm for eliminating potential noise. In the context of the study, points with a high level of direct illumination are considered the most informative, since they are naturally related to the parameters of the optical properties model of the scene. The point selection algorithm analyzes the geometry of the scene from several camera angles to the objects of the scene. This approach allows you to reduce the data set to directly visible points of the scene, and subsequently select the luminance ones from them. An algorithm for noise elimination was proposed and implemented for the correct analysis of scene images. This algorithm analyzes a point in the scene from different camera angles and performs interpolation in the area around this point. Thus, the idea of the image preprocessing algorithm is to select the most important points of the scene images, followed by selecting the luminance and eliminating noise. The results of the experiments show the correctness of the algorithm for selecting points with direct illumination and demonstrate the possibility of eliminating noise in synthetic images created with a small number of calculation iterations.
Fulltext:
PDF file (3095 kB)