Robust Denoising using Feature and Color Information

Supplemental material - comparisons and results

Rendering using UNIFORM sampling




All images were rendered at a 1024x1024 resolution using PBRT on a dual 6-core XEON system at 2.30GHz, with 12 rendering threads. In addition, the filtering step of our method is performed on a NVIDIA GeForce GTX TITAN. We compare renderings obtained using our method (OUR) with non-filtered renderings obtained using uniform low-discrepancy sampling (LD) and renderings obtained using the approach proposed by Li et al. (SURE, see paper for reference). For each image, we provide the average sample count per pixel (spp), the rendering time in seconds (s), and the relative MSE value. To account for the overhead of each method, all renderings of a given scene were obtained using an approximately equal rendering time. Indirect illumination is rendered using the path tracing integrator of PBRT. For the SURE method, we used the implementation provided by the authors, which runs on the CPU, using 12 threads.

We also provide the weights of the three candidate filters for our method. The weights of the FIRST, SECOND, and THIRD candidate filters are encoded using, respectively, the red, green, and blue channels.

We present comparisons to the method of Li et al. using adaptive rendering here: LWC12.
We present comparisons to the method of Dammertz et al. and Bauszat et al. using uniform rendering here: DSHL10, BEM11.
We present comparisons to the method of Kalantari and Sen, and Rousselle et al. using adaptive rendering here: KS13, RKZ12.
We present videos produced using spatial and space-time filtering here: videos.

Conference

area lighting, single bounce indirect illumination

neighborhood: r=20

conference

LD
100s; 152 spp


MSE: 207.719E-3

OUR
104s; 128 spp


MSE: 1.696E-3

SURE
104s; 128 spp


MSE: 4.560E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
256000 spp

Sanmiguel

environment lighting, single bounce indirect illumination

neighborhood: r=10

sanmiguel20

LD
55s; 36 spp


MSE: 283.355E-3

OUR
57s; 32 spp


MSE: 28.352E-3

SURE
67s; 32 spp


MSE: 43.279E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
64000 spp

Sibenik

area lighting, depth of field, single bounce indirect illumination

neighborhood: r=10

sibenik

LD
24s; 25 spp


MSE: 58.983E-3

OUR
25s; 16 spp


MSE: 1.321E-3

SURE
34s; 16 spp


MSE: 2.065E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
32000 spp

Teapot-metal

environment lighting, two bounces indirect illumination

neighborhood: r=10

teapot_metal

LD
19s; 33 spp


MSE: 178.682E-3

OUR
19s; 16 spp


MSE: 63.961E-3

SURE
29s; 16 spp


MSE: 89.230E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
4000 spp

Dragonfog

area lighting, depth of field, single bounce indirect illumination, single scattering participating media

neighborhood: r=20

dragonfog

LD
62s; 51 spp


MSE: 45.960E-3

OUR
59s; 32 spp


MSE: 1.556E-3

SURE
58s; 32 spp


MSE: 2.351E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
32000 spp

Plants-dusk

environment lighting, depth of field

neighborhood: r=10

plants_dusk

LD
42s; 19 spp


MSE: 2.274E-3

OUR
41s; 16 spp


MSE: 0.521E-3

SURE
53s; 16 spp


MSE: 0.698E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
4000 spp

Rings

area lighting, one bounce indirect illumination, caustics

neighborhood: r=20

rings

LD
32 spp


MSE: 164.414E-3

OUR
32 spp


MSE: 5.055E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
14400 spp

Sanmiguel 256 spp

environment lighting, single bounce indirect illumination

neighborhood: r=10

hq_sanmiguel20

LD
256 spp


MSE: 34.345E-3

OUR
256 spp


MSE: 7.852E-3

SURE
256 spp


MSE: 10.439E-3

OUR
CANDIDATES


WEIGHTS

REFERENCE
64000 spp

This webpage was built using the code found on: http://www.cs.huji.ac.il/~raananf/projects/lss_upscale/sup_images/index.html.