Hierarchical Image‐Space Radiosity for Interactive Global Illumination

Abstract

We introduce image-space radiosity and a hierarchical variant as a method for interactively approximating diffuse indirect illumination in fully dynamic scenes. As oft observed, diffuse indirect illumination contains mainly low-frequency details that do not require independent computations at every pixel. Prior work leverages this to reduce computation costs by clustering and caching samples in world or object space. This often involves scene preprocessing, complex data structures for caching, or wasted computations outside the view frustum. We instead propose clustering computations in image space, allowing the use of cheap hardware mipmapping and implicit quadtrees to allow coarser illumination computations. We build on a recently introduced multiresolution splatting technique combined with an image-space lightcut algorithm to intelligently choose virtual point lights for an interactive, one-bounce instant radiosity solution. Intelligently selecting point lights from our reflective shadow map enables temporally coherent illumination similar to results using more than 4096 regularly-sampled VPLs.

Direct illumination from dynamic area lights with visibility

“…In the future, we plan to adapt recent multiresolution techniques [Nichols et al 2009] to this problem, further improving performance. These techniques also smooth illumination samples during upsampling, which will filter out some of the high-frequency artifacts that can be seen in Figure 1.…”

Instant indirect illumination for dynamic mixed reality scenes

“…This reduces the total overdraw, which is clearly the bottleneck of the splatting approach. Later Nichols et al extend their approach to benefit from stencil buffering [27] and thus being able to perform the indirect lighting faster. Our splatting approach benefits from both approaches, that of Dachsbachers and that of Nichols.…”

Real-time rendering of deformable heterogeneous translucent objects using multiresolution splatting

“…Dachsbacher et al [4] employ reflective shadow maps to create VPLs to model single-bounce indirect lighting and gather the contributions via splatting [5]. Nichols et al extend on this by clustering VPLs [13] and hierarchical splatting [11,12]. While the proposed method also employs similar hierarchical data structures, key differences are the sampling and splatting strategies used to accommodate for differences in the density of the incident radiance between both: VPLs are sparsely distributed through the scene and their contribution covers the whole screen; conversely, in the case of heterogeneous subsurface scattering, the density of the irradiance samples is much higher, and their contribution covers a local screen region.…”

“…To efficiently detect geometrical variations (needed in step 2 in Fig. 2), we also construct an additional mipmap, similar to Nichols et al [13], to record the minimum and maximum depth values covered by each pixel in the hierarchical irradiance buffer. We initialize the minimum and maximum depth value at the finest level with the corresponding depth value at the finest irradiance level.…”

“…For each splatting buffer at all levels, we render the vertex array with the adaptive sampling implemented in vertex shaders. In particular, we follow the strategy of Nichols et al [13] to check the samples at all levels simultaneously (see also Algorithm 1). Next, the effective rendering range is computed in a vertex shader for all valid irradiance…”