Section: New Results

Complex Scenes

In order to render both efficiently and accurately ultra-detailed large scenes, this approach consists in developing representations and algorithms able to account compactly for the quantitative visual appearance of a regions of space projecting on screen at the size of a pixel.

Appearance pre-filtering

Participants : Guillaume Loubet, Neyret Fabrice.

We address the problem of constructing appearance-preserving level of details (LoDs) of complex 3D models such as trees and propose a hybrid method that combines the strength of mesh and volume representations. Our main idea is to separate macroscopic (i.e. larger than the target spatial resolution) and microscopic (sub-resolution) surfaces at each scale and to treat them differently, because meshes are very efficient at representing macroscopic surfaces while sub-resolution geometry benefit from volumetric approximations. We introduce a new algorithm based on mesh analysis that detects the macroscopic surfaces of a 3D model at a given resolution. We simplify these surfaces with edge collapses and provide a method for pre-filtering their BRDFs parameters. To approximate microscopic details, we use a heterogeneous microflake participating medium and provide a new artifact-free voxelization algorithm that preserves local occlusion. Thanks to our macroscopic surface analysis, our algorithm is fully automatic and can generate seamless LoDs at arbitrarily coarse resolutions for a wide range of 3D models. We validated our method on highly complex geometry and show that appearance is consistent across scales while memory usage and loading times are drastically reduced (see Figure 10). This work has been accepted at EG2017 [4].

Figure 10. A weeping willow 3D model pre-filtered with our method. Our LoDs use meshes for representing macroscopic surfaces and a volumetric representation to approximate sub-resolution geometry. This approach allows for accurate preservation of the appearance of complex geometry across scales while memory usage is drastic reduced. These images have been rendered with 256spp and a thin lense camera model in Mitsuba (http://www.mitsuba-renderer.org/).

Appearance pre-filtering of self-shadowing and anisotropic occlusion

Participants : Guillaume Loubet, Neyret Fabrice.

This year, we addressed the problem of representing the effect of internal self-shadowing in elements about to be filtered out at a given LOD, in the scope of volume of voxels containing density and phase-function (represented by a microflakes).

Naïve linear methods for downsampling high resolution microflake volumes often produce inaccurate results, especially when input voxels are very opaque. Preserving correct appearance at all resolutions requires taking into account inter- and intravoxel self-shadowing effects (see Figure 11). We introduce a new microflake model whose parameters characterize self-shadowing effects at the microscopic scale. We provide an anisotropic self-shadowing function and a microflake distribution for which scattering coefficients and phase functions of our model have closed-form expressions. We use this model in a new downsampling approach in which scattering parameters are computed from local estimations of self-shadowing in the input volume. Unlike previous work, our method handles datasets with spatially varying scattering parameters, semi-transparent volumes and datasets with intricate silhouettes. We show that our method generates LoDs with correct transparency and consistent appearance through scales for a wide range of challenging datasets, allowing for huge memory savings and efficient distant rendering without loss of quality. This work has been accepted at EG2018.

Figure 11. Comparison between naïve downsampling of microflake volumes and our method ("Aniso"). Naïve dowsampling of volumes with dense voxels often lead to inaccurate results due to the loss of inter- and intra-voxel self-shadowing effects. Our method is based on a new participating medium model and on local estimations of self-shadowing. It generates LoDs with correct transparency and consistent appearance through scales. Rendered with volume path tracing in Mitsuba (http://www.mitsuba-renderer.org/): the trunk of the cedar is a mesh.