RGL
Realistic Graphics Lab
EPFL Logo

A Learned Shape-Adaptive Subsurface Scattering Model

In Transactions on Graphics (Proceedings of SIGGRAPH 2019)

Ren­der­ing of trans­lu­cent soap blocks with sig­ni­fic­ant an­iso­tropy (g = 0.9). (a) Di­pole dif­fu­sion-based mod­els such as Photon Beam Dif­fu­sion [Habel et al. 2013] yield over­all flat ap­pear­ance due to their in­tern­al as­sump­tion of iso­tropy and plane-par­al­lel light trans­port. (b) Our learned sub­sur­face scat­ter­ing mod­el ad­apts to both geo­metry and an­iso­tropy, pro­du­cing more real­ist­ic ap­pear­ance and lower er­ror com­pared to a path-traced ref­er­ence. Ac­count­ing for the geo­metry leads to in­creased scat­ter­ing around the sil­hou­ette and over­all high­er con­trast in re­gions with geo­met­ric de­tail. Our meth­od has con­stant sampling weights, hence ren­der­ings con­verge with few­er samples. Please see the sup­ple­ment­al ma­ter­i­al for an in­ter­act­ive ver­sion of this fig­ure in­clud­ing er­ror maps.

Abstract

Sub­sur­face scat­ter­ing, in which light re­fracts in­to a trans­lu­cent ma­ter­i­al to in­ter­act with its in­teri­or, is the dom­in­ant mode of light trans­port in many types of or­gan­ic ma­ter­i­als. Ac­count­ing for this phe­nomen­on is thus cru­cial for visu­al real­ism, but ex­pli­cit sim­u­la­tion of the com­plex in­tern­al scat­ter­ing pro­cess is of­ten too costly. BSSRDF mod­els based on ana­lyt­ic trans­port solu­tions are sig­ni­fic­antly more ef­fi­cient but im­pose severe as­sump­tions that are al­most al­ways vi­ol­ated, e.g. planar geo­metry, iso­tropy, low ab­sorp­tion, and spa­tio-dir­ec­tion­al separ­ab­il­ity. The res­ult­ing dis­crep­an­cies between mod­el and us­age lead to ob­jec­tion­able er­rors in ren­der­ings, par­tic­u­larly near geo­met­ric fea­tures that vi­ol­ate planar­ity.

This art­icle in­tro­duces a new shape-ad­apt­ive BSSRDF mod­el that re­tains the ef­fi­ciency of pri­or ana­lyt­ic meth­ods while greatly im­prov­ing over­all ac­cur­acy. Our ap­proach is based on a con­di­tion­al vari­ation­al au­toen­coder, which learns to sample from a ref­er­ence dis­tri­bu­tion pro­duced by a brute-force volu­met­ric path tracer. In con­trast to the path tracer, our au­toen­coder dir­ectly samples out­go­ing loc­a­tions on the ob­ject sur­face, by­passing a po­ten­tially lengthy in­tern­al scat­ter­ing pro­cess.

The dis­tri­bu­tion is con­di­tion­al on both ma­ter­i­al prop­er­ties and a set of fea­tures char­ac­ter­iz­ing geo­met­ric vari­ation in a neigh­bor­hood of the in­cid­ent loc­a­tion. We use a low-or­der poly­no­mi­al to mod­el the loc­al geo­metry as an im­pli­citly defined sur­face, cap­tur­ing curvature, thick­ness, corners, as well as cyl­indric­al and tor­oid­al re­gions. We present sev­er­al ex­amples of ob­jects with chal­len­ging me­di­um para­met­ers and com­plex geo­metry and com­pare to ground truth sim­u­la­tions and pri­or work.