Evaluation of classic colour constancy algorithms on spectrally rendered ground-thought.

TY - CONF

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

AU - Toscani, Matteo

AU - Chen, Tao

AU - Guarnera, Claudio

PY - 2024/1/3

Y1 - 2024/1/3

N2 - The small number of available spectral images imposes a significant limit to colour science. We used computer graphics to overcome the problem and spectrally render naturalistic images, to investigate performance of classic colour constancy (CC) algorithms: 1) gray-world, 2) white-patch, 3) grayedge, 4) shades-of-gray, 5) gamut-mapping. We randomized the virtual observer's viewpoint and surface spectral reflectance in five indoor settings, resulting in a total of 4096 scenes. Instead of only comparing the illuminant estimated by the algorithms with the one inferred from a white surface embedded in each scene, we could evaluate CC algorithms based on the whole scene as rendered under a neutral illuminant. Performance of CC algorithms substantially differed between the two evaluation methods, indicating the importance of the whole-scene ground-truth. We selected a subsample of scenes and performed an online perceptual experiment to measure human CC and compare with predictions of the algorithms. Participants engaged in an asymmetric colour-matching task with a white sphere rendered at the centre of each scene. Human CC was better correlated with all algorithms’ performance when evaluated based on the whole scene. Furthermore, the arguably simplest algorithm (gray-world) best predicted human performance, suggesting that CC may be simpler that we thought.

AB - The small number of available spectral images imposes a significant limit to colour science. We used computer graphics to overcome the problem and spectrally render naturalistic images, to investigate performance of classic colour constancy (CC) algorithms: 1) gray-world, 2) white-patch, 3) grayedge, 4) shades-of-gray, 5) gamut-mapping. We randomized the virtual observer's viewpoint and surface spectral reflectance in five indoor settings, resulting in a total of 4096 scenes. Instead of only comparing the illuminant estimated by the algorithms with the one inferred from a white surface embedded in each scene, we could evaluate CC algorithms based on the whole scene as rendered under a neutral illuminant. Performance of CC algorithms substantially differed between the two evaluation methods, indicating the importance of the whole-scene ground-truth. We selected a subsample of scenes and performed an online perceptual experiment to measure human CC and compare with predictions of the algorithms. Participants engaged in an asymmetric colour-matching task with a white sphere rendered at the centre of each scene. Human CC was better correlated with all algorithms’ performance when evaluated based on the whole scene. Furthermore, the arguably simplest algorithm (gray-world) best predicted human performance, suggesting that CC may be simpler that we thought.

M3 - Abstract

T2 - Experimental Psychology Society meeting

Y2 - 3 January 2024

ER -