BCD Additional Materials
63
0
0
Fulltekst
(2) Supplementary results: compared methods For each scene, we compare: ARNLMF (Adaptive rendering with non-local means filtering) [Rousselle2012] RDFCI (Robust Denoising using Feature and Color Information) [Rousselle2013] WLR (Weighted Local Regression) [Moon2014] LBF (Learning Based Filtering) [Kalantari2015] RHF (Ray Histogram Fusion) [Delbracio2014] ours (Bayesian Collaborative Denoising) Note that ARNLMF and WLR are using adaptive sampling while ours performs a pure denoising and does not require triggering additional samples from the renderer.. BCD Additional Materials. 2 / 63.
(3) Supplementary results: timings. All computations are performed using a 4-core Intel Xeon processor at 2.80 GHz, and are always taking advantage of these 4 cores through CPU-parallelization. ARNLMF, RDFCI, WLR and LBF additionally use CUDA for GPU-parallelization (for filtering, not for rendering). All implementations use PBRT as rendering engine. ARNLMF, RDFCI, WLR and LBF are coded as PBRT extensions, and perform rendering, filtering and sometimes adaptive sampling in the same executable. RHF and ours use simple customizations of PBRT to produce additional inputs, then perform the filtering in a separate program. Hence, for timings of RHF and ours, we measure the filtering time and add it to the one that was needed to produce the input “Simple PBRT rendering”.. BCD Additional Materials. 3 / 63.
(4) Supplementary results: denoising measure. For each method, we measure the relative Mean Square Error (relMSE) and Structural Similarity (ssim) between the result and the groundtruth (generated with a much higher number of samples per pixel). A good denoising has a low relMSE (minimum 0) and a high ssim (maximum 1). All comparisons are equal-sample. Even the methods that perform adaptive sampling target the same average number of samples per pixel (spp). Note that some of the methods, like LBF, are able to provide nice results with a very low number of spp. But here, we assume a more production-like context where we can afford a reasonably large amount of samples per pixel and seek very high quality results.. BCD Additional Materials. 4 / 63.
(5) Supplementary results: scenes SanMiguelDof (256spp) Relatively complex scene with depth of field SponzaDir (256spp) Quite dark, needs multiple bounces to reach the only (directional) light. TeapotMetal (64spp) Simple scene with specular materials RingsPhotonMap (128spp) Several caustics, uses Photon Mapping for rendering (all other scenes are using standard path tracing). SponzaKilleroo (512spp) Same as SponzaDir, but with an additional golden object that cause caustics and many spikes. SibenikCar (256spp) Highly specular materials (car and ground). SibenikCrash (256spp) Same as SibenikCar, but with motion blur on the car. All images have resolution 1024 × 1024 BCD Additional Materials. 5 / 63.
(6) SanMiguelDof — Simple PBRT rendering. (256 spp). Time: 2732 s ssim: 0.5913 — relMSE: 2065 × 10−4. BCD Additional Materials. 6 / 63.
(7) SanMiguelDof — ARNLMF Time (adaptive sampling and reconstruction): 3313 s ssim: 0.9318 — relMSE: 90 × 10−4. BCD Additional Materials. 7 / 63.
(8) SanMiguelDof — RDFCI Time (rendering + filtering): 2803 s ssim: 0.9639 — relMSE: 70 × 10−4. BCD Additional Materials. 8 / 63.
(9) SanMiguelDof — WLR Time (adaptive sampling and reconstruction): 3506 s ssim: 0.9585 — relMSE: 68 × 10−4. BCD Additional Materials. 9 / 63.
(10) SanMiguelDof — LBF Time (rendering + filtering): 3397 s ssim: 0.9429 — relMSE: 95 × 10−4. BCD Additional Materials. 10 / 63.
(11) SanMiguelDof — RHF Time (rendering + filtering): 2732 + 232 = 2964 s ssim: 0.9324 — relMSE: 122 × 10−4. BCD Additional Materials. 11 / 63.
(12) SanMiguelDof — ours Time (rendering + filtering): 2732 + 37 = 2769 s ssim: 0.9350 — relMSE: 91 × 10−4. BCD Additional Materials. 12 / 63.
(13) SanMiguelDof — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 13 / 63.
(14) SponzaDir — Simple PBRT rendering. (256 spp). Time: 966 s ssim: 0.8191 — relMSE: 1631 × 10−4. BCD Additional Materials. 14 / 63.
(15) SponzaDir — ARNLMF Time (adaptive sampling and reconstruction): 1114 s ssim: 0.9386 — relMSE: 398 × 10−4. BCD Additional Materials. 15 / 63.
(16) SponzaDir — RDFCI Time (rendering + filtering): 964 s ssim: 0.9728 — relMSE: 157 × 10−4. BCD Additional Materials. 16 / 63.
(17) SponzaDir — WLR Time (adaptive sampling and reconstruction): 1175 s ssim: 0.9707 — relMSE: 175 × 10−4. BCD Additional Materials. 17 / 63.
(18) SponzaDir — LBF Time (rendering + filtering): 1286 s ssim: 0.9537 — relMSE: 269 × 10−4. BCD Additional Materials. 18 / 63.
(19) SponzaDir — RHF Time (rendering + filtering): 966 + 270 = 1236 s ssim: 0.9000 — relMSE: 587 × 10−4. BCD Additional Materials. 19 / 63.
(20) SponzaDir — ours Time (rendering + filtering): 966 + 22 = 988 s ssim: 0.9426 — relMSE: 276 × 10−4. BCD Additional Materials. 20 / 63.
(21) SponzaDir — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 21 / 63.
(22) TeapotMetal — Simple PBRT rendering. (64 spp). Time: 186 s ssim: 0.6422 — relMSE: 1802 × 10−4. BCD Additional Materials. 22 / 63.
(23) TeapotMetal — ARNLMF Time (adaptive sampling and reconstruction): 222 s ssim: 0.7579 — relMSE: 1488 × 10−4. BCD Additional Materials. 23 / 63.
(24) TeapotMetal — RDFCI Time (rendering + filtering): 165 s ssim: 0.8671 — relMSE: 641 × 10−4. BCD Additional Materials. 24 / 63.
(25) TeapotMetal — WLR Time (adaptive sampling and reconstruction): 255 s ssim: 0.8431 — relMSE: 795 × 10−4. BCD Additional Materials. 25 / 63.
(26) TeapotMetal — LBF Time (rendering + filtering): 247 s ssim: 0.8232 — relMSE: 764 × 10−4. BCD Additional Materials. 26 / 63.
(27) TeapotMetal — RHF Time (rendering + filtering): 186 + 259 = 445 s ssim: 0.7763 — relMSE: 1164 × 10−4. BCD Additional Materials. 27 / 63.
(28) TeapotMetal — ours Time (rendering + filtering): 186 + 41 = 227 s ssim: 0.8160 — relMSE: 841 × 10−4. BCD Additional Materials. 28 / 63.
(29) TeapotMetal — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 29 / 63.
(30) RingsPhotonMap — Simple PBRT rendering. (128 spp). Time: 16904 s ssim: 0.5358 — relMSE: 1413 × 10−4. BCD Additional Materials. 30 / 63.
(31) RingsPhotonMap — ARNLMF Time (adaptive sampling and reconstruction): 16918 s ssim: 0.8271 — relMSE: 1148 × 10−4. BCD Additional Materials. 31 / 63.
(32) RingsPhotonMap — RDFCI Time (rendering + filtering): 11084 s ssim: 0.8106 — relMSE: 1214 × 10−4. BCD Additional Materials. 32 / 63.
(33) RingsPhotonMap — WLR Time (adaptive sampling and reconstruction): 14433 s ssim: 0.8938 — relMSE: 925 × 10−4. BCD Additional Materials. 33 / 63.
(34) RingsPhotonMap — LBF Time (rendering + filtering): 17758 s ssim: 0.7718 — relMSE: 2107 × 10−4. BCD Additional Materials. 34 / 63.
(35) RingsPhotonMap — RHF Time (rendering + filtering): 16904 + 286 = 17190 s ssim: 0.8460 — relMSE: 886 × 10−4. BCD Additional Materials. 35 / 63.
(36) RingsPhotonMap — ours Time (rendering + filtering): 16904 + 56 = 16960 s ssim: 0.8433 — relMSE: 498 × 10−4. BCD Additional Materials. 36 / 63.
(37) RingsPhotonMap — Groundtruth 4096 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 37 / 63.
(38) SponzaKilleroo — Simple PBRT rendering. (512 spp). Time: 2109 s ssim: 0.7830 — relMSE: 7125 × 10−4. BCD Additional Materials. 38 / 63.
(39) SponzaKilleroo — ARNLMF Time (adaptive sampling and reconstruction): 2373 s ssim: 0.9459 — relMSE: 430 × 10−4. BCD Additional Materials. 39 / 63.
(40) SponzaKilleroo — RDFCI Time (rendering + filtering): 2150 s ssim: 0.9652 — relMSE: 245 × 10−4. BCD Additional Materials. 40 / 63.
(41) SponzaKilleroo — WLR Time (adaptive sampling and reconstruction): 2374 s ssim: 0.9671 — relMSE: 260 × 10−4. BCD Additional Materials. 41 / 63.
(42) SponzaKilleroo — LBF Time (rendering + filtering): 2735 s ssim: 0.9572 — relMSE: 257 × 10−4. BCD Additional Materials. 42 / 63.
(43) SponzaKilleroo — RHF Time (rendering + filtering): 2109 + 276 = 2385 s ssim: 0.9300 — relMSE: 489 × 10−4. BCD Additional Materials. 43 / 63.
(44) SponzaKilleroo — ours Time (rendering + filtering): 2109 + 28 = 2137 s ssim: 0.9520 — relMSE: 287 × 10−4. BCD Additional Materials. 44 / 63.
(45) SponzaKilleroo — ours with spike-removal prefiltering Time (rendering + filtering): 2109 + 28 = 2137 s ssim: 0.9523 — relMSE: 267 × 10−4. BCD Additional Materials. 45 / 63.
(46) SponzaKilleroo — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 46 / 63.
(47) SibenikCar — Simple PBRT rendering. (256 spp). Time: 2495 s ssim: 0.6338 — relMSE: 995 × 10−4. BCD Additional Materials. 47 / 63.
(48) SibenikCar — ARNLMF Time (adaptive sampling and reconstruction): 2703 s ssim: 0.9541 — relMSE: 529 × 10−4. BCD Additional Materials. 48 / 63.
(49) SibenikCar — RDFCI Time (rendering + filtering): 2419 s ssim: 0.9694 — relMSE: 53 × 10−4. BCD Additional Materials. 49 / 63.
(50) SibenikCar — WLR Time (adaptive sampling and reconstruction): 2701 s ssim: 0.9599 — relMSE: 102 × 10−4. BCD Additional Materials. 50 / 63.
(51) SibenikCar — LBF Time (rendering + filtering): 2911 s ssim: 0.9495 — relMSE: 179 × 10−4. BCD Additional Materials. 51 / 63.
(52) SibenikCar — RHF Time (rendering + filtering): 2495 + 258 = 2753 s ssim: 0.9637 — relMSE: 363 × 10−4. BCD Additional Materials. 52 / 63.
(53) SibenikCar — ours Time (rendering + filtering): 2495 + 33 = 2528 s ssim: 0.9602 — relMSE: 95 × 10−4. BCD Additional Materials. 53 / 63.
(54) SibenikCar — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 54 / 63.
(55) SibenikCrash — Simple PBRT rendering. (256 spp). Time: 3925 s ssim: 0.6093 — relMSE: 1038 × 10−4. BCD Additional Materials. 55 / 63.
(56) SibenikCrash — ARNLMF Time (adaptive sampling and reconstruction): 4212 s ssim: 0.9576 — relMSE: 45 × 10−4. BCD Additional Materials. 56 / 63.
(57) SibenikCrash — RDFCI Time (rendering + filtering): 3557 s ssim: 0.9712 — relMSE: 32 × 10−4. BCD Additional Materials. 57 / 63.
(58) SibenikCrash — WLR Time (adaptive sampling and reconstruction): 4082 s ssim: 0.9656 — relMSE: 46 × 10−4. BCD Additional Materials. 58 / 63.
(59) SibenikCrash — LBF Time (rendering + filtering): 4329 s ssim: 0.9585 — relMSE: 64 × 10−4. BCD Additional Materials. 59 / 63.
(60) SibenikCrash — RHF Time (rendering + filtering): 3925 + 255 = 4180 s ssim: 0.9662 — relMSE: 38 × 10−4. BCD Additional Materials. 60 / 63.
(61) SibenikCrash — ours Time (rendering + filtering): 3925 + 29 = 3954 s ssim: 0.9625 — relMSE: 42 × 10−4. BCD Additional Materials. 61 / 63.
(62) SibenikCrash — Groundtruth 65536 spp ssim: 1 — relMSE: 0. BCD Additional Materials. 62 / 63.
(63) Comparison diagrams. BCD Additional Materials. 63 / 63.
(64)
RELATERTE DOKUMENTER
We compute seismograms for 3D hetero- geneous acoustic media using this technique and compare these seismograms with seismograms computed using two other modeling
Traditional direct volume rendering techniques consist in in- tegrating opacity and color values along a viewing ray by using a simple light transport model, called the volume
The columns from left to right: name of the test scene; resolution; time taken by direct illumination including shadow map rendering and percentage-closer filtering; rendering
We compared our PMC-ER equal deposition algorithm with the energy redistribution path tracing (ERPT) algorithm on the Cornell Box scene, a dragon scene, and a complex room scene
• Local AND global illumination model needed – Path propagation - global model – BRDF evaluation - local model. •
The primary goal of this dissertation is to develop robust and general noise reduction methods for interactive progressive stochastic ray tracing algo- rithms.. Robust means the
To compare the results of the Tolerance algorithm with hu- man perception, first we calculated how often the subjects classified each stroke as good, average or poor, using the
Figure 3: The guide image and optimization result of the Lucy scene, and rendering results with foliage shadow effects from different viewing angles.... (a) Scene setup for the
