290 likes | 453 Views
Perceptual Influence of Approximate Visibility in Indirect Illumination. Insu Yu 27 May 2010. ACM Transactions on Applied Perception (Presented at APGV 2009). Introduction. Can you see difference ?
E N D
Perceptual Influence of Approximate Visibility in Indirect Illumination Insu Yu 27 May 2010 ACM Transactions on Applied Perception (Presented at APGV 2009)
Introduction • Can you see difference ? Traditionally GI (Path tracing, photon mapping, ray-tracing) uses accurate visibility (ray casting)
Motivation Direct Indirect Direct + Indirect • Indirectillum. is perceptually important in GI but high rendering cost • Low frequency nature in real world(Smooth gradation) • Visibility determination is most expensive • Intersecting rays with all polygons • Approximated Visibility for efficient GI?
Previous works (Approx. Visibility) • Radiosity(Sillion 95): blur out small features • Lightcuts(Walter05) : grouping VLPs • Interactive GI for dynamic(Dachsbacher05) : indirect illum is neglected • GPU-based indirect illum(Bunnel05): a hierarchical link structure • Ambient occlusion (Zhukov98) • Hierarchical radiosity(Dachsbacher07) for IGI • Imperfect Shadow(Ritscehl08b) • Approx Visibility used in RT but perceptual impact not formally studied • No distinction of direct / indirect illumination
Overview • Visibility Approximations (in our study) • Imperfect Visibility(Ritschel 2008) • Ambient Occlusion(Zhukov1998) • Direction Ambient Occlusion(Sloan 09. Ritschel 09) • Evaluate Perceptual influence of approximated visibility on scene with indirect illumination • Which visibility approx are perceptually acceptable • Scope • Only interested in indirect illumination • Direct illumination uses accurate visibility • Evaluate with Instant radiosity method
Rendering Equation (Visibility ?) • Rendering Equation • Reflection operator ‘K’, Geometry operator ‘G’, visibility ‘V’ • Operator form • KG < 1
Path dependant notation • Neumann expansion • Path notation • G0 Direct lighting - use accurate visibility (Ga) • G1 first bounce • G2 light is reflected the second time • Superscript dots: path-length • Subscript: bounce number
Path dependant notation IR (VPLs) shoot photons from light sources deposit on every bounce treat photons as point lights Direct illumination Indirect first bounce Shadow Ray
Approximations - IMP • Imperfect Visibility (Gimp) • Ritschel08 used to speed up instant radiosity • Randomly setting N% of visibility to either 0 or 1 • Introduce noise • Last bounce is approximated X X Accurate Imperfect
Approximations – AO • Ambient Occlusion(Gao) • Produce smooth visibility • The percentage of ‘visible sky’ -scalar value • visibility from ‘x’ in all directions • User defined radius ‘r’ Accurate Ambient Occlusion
Approximations – DAO • Directional Ambient Occlusion(Gdao) • Add directional component to AO • Partial correct and errors • 5th order Spherical Harmonics to represent directional visibility Accurate Dir. Ambient Occlusion
Approximations – Cont’ • No Visibility (Gno) • No visibility indirect illumination for Interactive GI • Validate whether no visibility is useful approximation • V(x,y) = 1
Perceptual Influence Study • Goal • Evaluate the influence of visibility approximations • Carry out a series of psychophysical experiments • How perceptually similar to a reference • Paired comparison • Visibility approximations appear realistic (perceived realism) • Ranking Order
Stimuli • Chosen parameter sets to speed up for real-time apps (IR, lightcuts, path-tracing) • l is path length • Accurate Visibility • IMP: (case 1~3) • 25% 50% 75% visibility corruptions • AO,DAO (case 4-6, 7-9) • r = 0.05, 0.1, 0.2 radius of scene diameter • No visibility (case 10) Accurate Visibility for Direct lighting
Test Scenes Arches Tea house LivingroomSponza • Five seconds video instead of static images to take into account of temporal artifacts • A full range of scenes Arches : Fast light moving + strong direct lighting Tea house: Slow light moving + dominant Indirect Livingroom: Camera moving + dominant Direct Lighting Sponza: Fast Camera moving + strong indirect shadows
Test Scenes (cont’) Arches Tea house LivingroomSponza • Rendered using Instant Radiosity(IR) • Four indirect illuminations • High number of VPLs to avoid artifaces • 640x480 resolutions and gamma corrected • Rendered in a PC cluster (1-4 hrs per image)
Experiment- Paired comparison • Paired comparison plus category (Scheffe52) • Quantify perceptual similarity to the reference • How similar to the reference (pair of videos) • Five-point scoring scale • Assign 1(not similar), 2(slightly similar), 3(moderately similiar), 4(very much similar), 5 (extremely similar) • Category rating + pair comparison
Experiment - Ordinal rank order • Ordinal rank order (Bartleson84) • Determine the perceived realism • Can be quickly performed than complete pair-wise comparison • Intuitive user interface (videos in a row is shown) • Rank the videos in order from highest to lowest by perceived realism
Experiment Procedure • Procedure • Two sessions on different days • A training session was given • conducted in a controlled environment • Paired comparison • 40 estimates (4 scenes x 10 approx) took 15mins • 14 subjects for paired comparisons • Ranking order • 11 videos(reference + 10 approximations) • Rank 4 different sets, sorting 11 videos (25-35mins) • 18 subjects for ranking experiment • Pan/Zoom, Drag & Drop, Pause functions
Results and Analysis • Perceptual Scales: • Five-point scores were scaled using Law of Categorical Judgment Torgerson(58) and Thurstonia(27) • Category boundaries • Estimates of the category boundaries • The scale values can be related to the original categories
Results (Pair comparison) AO & DAO: slightly similar (radius >= 0.2) – Large Radius IMP: very much /moderately similar (accepted wide range of scenes) AO: very much /moderately similar (radius < 0.1) DAO: very much /moderately similar (radius < 0.1) Direct illumination dominant No Vis: moderately similar Arches Tea house Living room Sponza
Results (Ranking) No Vis: Ranked higher than worst AO, DAO Ref, IMP: ranked equally very realistic AO & DAO(0.05, 0.1): ranked generally realistic AO & DAO(0.2): ranked less realistic Arches Tea house Living room Sponza
Results (Overall) All IMP perceived very high realism. AO,DAO: Large radius perceived significantly less realistic AO,DAO(r=0.2) less realistic than ‘no visibility’ Correlation between similarity and perceived realism
Discussion • Visibility approximations can be used in GI maintaining appearance is perceptually similar to ref • IMP ranked generally higher ‘perceived realism’ • Highly corrupted(random) is preferred to human eyes than inaccurate AO,DAO • Most visibility approx are ‘very much similar’ to the ref when direction illumination is dominant • Validates the use of visibility approximations
Thanks you Questions ?