Programmirovanie

The paper studies surface rendering methods based on ray tracing for representations based on signed distance functions. The main objects of interest were the rendering algorithm execution time, the amount of memory occupied, and the accuracy of the surface representation estimated by the render using the PSNR metric. Six different representations and four intersection search algorithms were analyzed. In all cases, a bounding volume hierarchy was used as an accelerating structure. The comparison revealed promising representations and algorithms and showed that distance functions in some cases are not inferior to polygonal models in speed, while they can win in terms of memory consumption and represent the surface with a good level of accuracy.

The use of differentiable rendering methods is an up-to-date solution to the problem of geometry reconstruction from a set of RGB images without using expensive equipment. The disadvantage of this class of methods is the possible distortions of the geometry that arise during optimization and high computational complexity. Modern differentiable rendering methods calculate and use two types of gradients: silhouette gradients and normal gradients. Most distortions arising in geometry optimization are caused by modifications of parameters associated with silhouette gradients. The paper considers the possibility of increasing the efficiency of geometry reconstruction methods based on the use of differentiable rendering by dividing the reconstruction process into two stages: initialization and optimization. The first stage of reconstruction involves the creation of a visual shell of the reconstructed object. This stage allows one to automate the process of selecting the original geometry and start the next stage under two conditions: the silhouettes of the object have already been reconstructed from all observation points and the topologies of the reconstructed and true objects are equivalent. The second stage comprises a geometry optimization cycle based on the fulfillment of the above conditions. This cycle consists of four steps: image rendering, loss function calculation, gradient calculation, and geometry optimization. Satisfying the condition of matching the contours of the original and reference geometry eliminates the need to use silhouette gradients. This solution significantly reduces the number of errors that occur during optimization, as well as reduces the computational complexity of the method by eliminating the calculation of the loss function, gradient calculation, and optimization of parameters associated with the silhouettes of objects. The testing and analysis of the results showed an increase in the accuracy of geometry reconstruction with a decrease in grid resolution and a decrease in the total running time of the method in comparison with similar methods, as well as an up to two-fold increase in the speed of optimization steps.

A way to improve the quality of magnetic resonance image processing using the Kolmogorov–Arnold networks for deep feature filtering in the convolutional neural network is studied. Recently proposed Kolmogorov–Arnold networks are inspired by the representation theorem of the same name from real analysis and approximation theory. It states that every multivariate continuous function on a compact set can be represented as a superposition of continuous single-variable functions. However, further gradient descent application imposes restrictions on the inner Kolmogorov functions to be at least differentiable, that’s why, in practice, they are searched in a linear span of B-Splines or some other differentiable basis functions. In this study we propose an adaptive method of basis functions selection by the model itself during training, mitigating the rule of thumb choice of that basis functions. The method is based on the attention mechanism, successfully used in state-of-the-art transformers. The proposed approach is tested on magnetic resonance images enhancement on IXI dataset and demonstrates the best average PSNR and TV over the synthetic testing dataset. Without loss of generality, the system of basis functions included: B-splines, Chebyshev polynomials and Hermite functions.

A key part of the most common ray tracing methods is the traversal/search for intersections with a hierarchical structure – BVH, which describes the scene geometry. This paper presents a comparative analysis of the performance of several BVH tree traversal methods on stationary and mobile graphics processors. We investigated BVH trees with varying depths and numbers of child nodes, implemented several stack-based traversal algorithms, and two different stackless traversal algorithms; we proposed our own stackless traversal variant, which is more performant than existing ones in some cases. We proposed our own compression method for BVH trees with 2 nodes, losing no more than 15% performance. We identified a common problem that occurs in almost all algorithms when implemented on graphics processors. We believe that our analysis will help developers of ray tracing hardware accelerators create more economical hardware solutions, not limited solely to ray tracing. More specifically, our experimental results suggest that up to a 5x speedup can be achieved by changing the L2 cache mechanism, and this has likely already been implemented in stationary GPUs with hardware-accelerated ray tracing, not only within the ray tracing mechanism itself but also in a more general context.

With the increase in GPU performance, it has become possible to visualize complex physical phenomena in real time using global lighting algorithms. One such approach is the use of virtual point light sources, in which the realism of images depends on the number of light sources. However, for a large number of light sources, early algorithms required the creation of a large number of shadow maps to check visibility under virtual spot lighting. Therefore, it was problematic to achieve high-quality real-time images until new methods were developed. The purpose of the presented work is to create a delayed rendering method for thousands of point light sources based on voxelized scenes in real time. In the first pass, a geometric sparse voxel octal tree is calculated. A geometric buffer is used that stores information about location, normal, and materials for direct and indirect lighting. Then, reflective shadow maps are generated and selected by significance, so as not to check each texel. Direct illumination is calculated using shadow maps, and for indirect illumination, a ray marching algorithm is used to check the visibility of point light sources. Alternating sampling is used to speed up calculations. As a result, using the proposed method, it is possible to create realistic images of scenes with global illumination in real time. Using a GPU, you can calculate thousands of point light sources in real time, and visualize fully dynamic scenes. However, glossy surfaces require a greater number of point light sources so that images without artifacts accurately reproduce the appearance of the material.