Evaluation of classic colour constancy algorithms on spectrally rendered ground-truth

TY - CONF

T1 - Evaluation of classic colour constancy algorithms on spectrally rendered ground-truth

AU - Toscani, Matteo

AU - Guarnera, Claudio

AU - Chen, Tao

PY - 2023/8/27

Y1 - 2023/8/27

N2 - The small number of available spectral images imposes a significant limitation to colour science. We used computer graphics techniques to spectrally render naturalistic images, used to investigate the performance of three classic colour constancy algorithms: 1) grey-world, 2) white-patch, 3) grey-edge.Rather than comparing the illuminant estimated by the algorithms with the one inferred from a white surface embedded in each scene, we evaluated existing colour-constancy algorithms based on the spectral images. For each of two indoor settings, we randomized the point of view of the virtual observer and thespectral reflectance of the surfaces, for a total of 50 random scenes per setting. We sampled random reflectances from a compact statistical model obtained by applying a Principal Component Analysis to spectral reflectance measures. We rendered each of these random scenes under 5 different illuminants, linearly spaced along the daylight locus, from blue (CIE x=0.2, y=0.18) to yellow (CIE x=0.6, y=0.37), for a total of 500 scenes. We evaluated the performances ofthe different algorithms by computing the angular error between 1) the groundtruth illuminant and the estimated illuminants, 2) the rgb colour of the pixels of each scene rendered under the equal-energy illuminant E and the colour of the pixels rendered under the other illuminants. We found that differences in performancebetween algorithms are massively reduced when evaluated on the whole scene rather than on the illuminated estimates. In particular, when using the white-patch or the grey-edge algorithms, the angular errors are respectively 89% and 87% then the grey-world algorithm one. Crucially, when we only considered the estimated illuminants, the angular error obtained with the white-patch or the grey-edge algorithms is much less (1% and 52%). Our, results highlight the importance of spectral ground-truth for colour constancy research, as well as the potential contribution of computer graphics to colour science.

AB - The small number of available spectral images imposes a significant limitation to colour science. We used computer graphics techniques to spectrally render naturalistic images, used to investigate the performance of three classic colour constancy algorithms: 1) grey-world, 2) white-patch, 3) grey-edge.Rather than comparing the illuminant estimated by the algorithms with the one inferred from a white surface embedded in each scene, we evaluated existing colour-constancy algorithms based on the spectral images. For each of two indoor settings, we randomized the point of view of the virtual observer and thespectral reflectance of the surfaces, for a total of 50 random scenes per setting. We sampled random reflectances from a compact statistical model obtained by applying a Principal Component Analysis to spectral reflectance measures. We rendered each of these random scenes under 5 different illuminants, linearly spaced along the daylight locus, from blue (CIE x=0.2, y=0.18) to yellow (CIE x=0.6, y=0.37), for a total of 500 scenes. We evaluated the performances ofthe different algorithms by computing the angular error between 1) the groundtruth illuminant and the estimated illuminants, 2) the rgb colour of the pixels of each scene rendered under the equal-energy illuminant E and the colour of the pixels rendered under the other illuminants. We found that differences in performancebetween algorithms are massively reduced when evaluated on the whole scene rather than on the illuminated estimates. In particular, when using the white-patch or the grey-edge algorithms, the angular errors are respectively 89% and 87% then the grey-world algorithm one. Crucially, when we only considered the estimated illuminants, the angular error obtained with the white-patch or the grey-edge algorithms is much less (1% and 52%). Our, results highlight the importance of spectral ground-truth for colour constancy research, as well as the potential contribution of computer graphics to colour science.

M3 - Abstract

T2 - EUROPEAN CONFERENCE ON VISUAL PERCEPTION

Y2 - 27 August 2023

ER -