Real vs Render: Colorimetric and Perceptual Accuracy using a Real and Rendered Cornell Box with HMD-VR

Walkom Gareth – KU Leuven

The Cornell box, consisting of gray walls with one red and one green wall, has been adopted in many computer graphic studies to show the interaction of light in computer renderings versus real scenes. Although the Cornell Box is used throughout studies in computer graphics, it is currently unknown how the colorimetric and perceptual accuracy of the Cornell Box renderings correspond to that of a real Cornell Box.
As DIALux is one of the most commonly used software packages by industry for simulating real lighting scenes, the Cornell Box will be modeled using DIALux and also the current state-of-the-art, spectral and physical based renderer (SPBR), Mitsuba. Comparisons of colorimetric and perceptual accuracy will be made with the real Cornell Box, a Cornell Box in DIALux, and a Cornell Box in Mitsuba.
A real Cornell Box will be built using the same characteristics of Cornell Boxes in literature, with the walls and objects layered with uniformly colored paper, and the spectral reflectance of the colored paper measured. Once assembled, the XYZ tristimulus maps of the Cornell Box will be acquired using a Luminance Camera and the spectral irradiance of the real Cornell Box will then be also measured, using a spectral irradiance meter.
A 3D model of the Cornell Box will be accurately created in DIALux evo using the same measurements as the real Cornell Box. DIALux evo allows Light-Intensity-Distribution (LID) (LDT files) to be easily added to a scene, accurately showing the same distribution of light as a real luminaire. The measured spectral reflectance data of the paper – assumed to have a Lambertian Bi-Directional-Reflectance-Distribution (BRDF) in DIALux – will be added to the objects accordingly in the DIALux scene to model the scenes as closely as possible.
Similar to in DIALux, a 3D model of the Cornell Box will be produced in Blender using the same measurements as the real box, and then rendered in Mitsuba. As Mitsuba does not have a function to include LDT files in a scene but is however open-source, a function that uses a lamp with a texture to portray a LID can be added. The BRDF of the paper shall be measured in a goniophotometer and modeled with a combination of a diffuse and a specular BRDF. The fitted parameters and the measured spectral reflectance will then be used as input to Mitsuba. The XYZ tristimulus values are then acquired from the same point of view and with the same field of view as the luminance camera in the real scene.
To evaluate the colorimetric accuracy of the real Cornell Box, the Cornell Box in DIALux, and the Cornell Box in Mitsuba, the measured XYZ tristimulus maps will be compared to analyze the luminance difference and CIE 1976 u’v’ chromaticity difference.
The rendered Cornell Box from Mitsuba will then also be visually evaluated using Head-Mounted Display Virtual Reality (HMD-VR) for perceptual accuracy in terms of perceived brightness, colorfulness, hue, and contrast in comparison with the real Cornell Box.