How well can categorical colour perception be maintained in natural environments with varying illuminants? To address this question, a colour-naming experiment was performed with colour-monitor images of natural scenes simulated under two different daylights of correlated colour temperature 6500 K and 25000 K. Images were obtained from a set of hyperspectral data to enable the accurate control of illuminant and reflectance spectra. Each scene contained a spherical test surface whose digitally manipulated spectral reflectance coincided with that of a sample drawn randomly from approximately 430 Munsell reflectances grouped into eight colour categories, namely, red, green, blue, yellow, pink, purple, brown, and orange. Observers had to name the colour of the test surface in each image, presented for 1 s, by pressing one of nine computer keys corresponding to the eight categorical colour names and a neutral category. Focal colours were estimated from the peaks of the smoothed distributions of observers' naming responses in the CIE 1976 (u′, v′) chromaticity diagram. The effect of the illuminant change was quantified by a focal colour constancy index, with values 0 and 1 corresponding to no constancy and perfect constancy. Average levels of focal colour constancy were close to those from traditional measures of colour constancy, but with variation across categories and surface lightness. For blue and purple surfaces levels approached 0.9. For many surface colours, colour naming seems to be robust under illuminant changes and may help to anchor non-categorical judgements of arbitrary surface colours in natural scenes.