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

The limited availability of spectral images poses a significant challenge to the field of colour science. To address this issue, we spectrally rendered naturalistic images, enabling us to investigate the performance of classic colour constancy algorithms, including <i>Grey-World</i>, <i>White-Patch</i>, <i>Grey-Edge</i>, <i>Shades-of-Grey</i>, and <i>Gamut-Mapping</i>. We generated 4,096 physically based rendered scenes under different coloured illuminations, including a spectrally neutral illumination. We evaluated each algorithm by (1) comparing the illuminant estimated by the algorithm with the actual illuminant used for rendering and (2) assessing the performance based on the entire scene rendered under the neutral illuminant. The <i>White-Patch</i> algorithm consistently performed relatively well, while <i>Gamut-Mapping</i> emerged as the top-performing algorithm when evaluating the whole scene. However, it exhibited poor performance in estimating the ground-truth illuminant. We conducted a perceptual experiment to measure human colour constancy across a representative selection of scenes from our database using an asymmetric colour matching task. The results indicated that predictions from the algorithms that performed best when evaluated on the whole scene - <i>white patch</i> and <i>gamut mapping</i> - best approximate human performance. Indeed, the function of colour constancy is to stabilise the colour of all surfaces in a scene, rather than to estimate the colour of the illumination.