When the brightest is not the best: illuminant estimation from the geometry of specular highlights

Abstract

Colour constancy supports stable object-colour perception across changes in illumination. Illuminant colour can be inferred from white surfaces or specular highlights, and many models adopt a "brightest is white" heuristic to identify illuminant colour. We tested an alternative: observers use structured changes in the proximal image to locate highlights even when they are not the brightest elements. Using computer-rendered scenes, we manipulated the reliability of two cues--brightest element and highlight geometry. Each scene showed a single textured sphere lit by multiple point lights with identical spectra. The sphere had uniform spectral reflectance, while a noise texture attenuated the reflectance by a variable scale factor. We tested three specularity levels (matte, low, mid). Observers viewed a 1.5 s animation and judged whether chromatic changes reflected an illuminant or material change. Performance for matte surfaces was near chance but improved sharply with increasing specularity. Observers outperformed an ideal observer relying solely on the brightest element. When highlights landed on darker texture regions, performance improved further, contrary to the brightest-element prediction. Phase scrambling, which disrupted highlight geometry while preserving global statistics, substantially reduced performance. Overall, observers do not rely solely on the brightest element; they exploit diffuse-specular regularities to resolve surface-illuminant ambiguity.