Studies on colored transparent objects have elucidated potential mechanisms, but these studies have mainly focused on flat filters overlaying flat backgrounds. While they have provided valuable insight, these studies have not captured all aspects of transparency, like caustics, specular reflections/highlights, and shadows. Here, we investigate color-matching experiments with curved transparent objects for different matching stimuli: a uniform patch and a flat filter. Two instructions were tested: simply match the color of the glass object and the test element (patch and flat filter) or match the color of the dye that was used to tint the transparent object (patch). Observers’ matches differed from the mean, the most frequent, and the most saturated color of the transparent stimuli, whereas the brightest regions captured the chromaticity, but not the lightness, of patch matches. We applied four models from flat filter studies: the convergence model, the ratios of either the means (RMC) or standard deviations (RSD) of cone excitations, and a robust ratio model. The original convergence model does not fully generalize but does not perform poorly, and with modifications, we find that curved transparent objects cause a convergence of filtered colors toward a point in color space, similar to flat filters. Considering that, the RMC and robust ratio models generalized more than the RSD, with the RMC performing best across the stimuli we tested. We conclude that the RMC is probably the strongest factor for determining the color. The RSD seems instead to be related to the perceived “clarity” of glass objects.
Observers readily answer the question, “What is the color of this transparent object?” regardless of whether the matching element is a uniform patch or a flat transparent filter. In both cases, their responses differ from the mean chromaticity of the Glaven, as well as its most saturated color and its most frequent color, while the White Point almost captures the uniform patch match but falls short of explaining the lightness of their settings. At least for a flat filter matching element, observer responses are the result of a color constancy-esque discounting operation. More accurately, we can say that the ratios of the mean cone excitations (RMC) between the filtered and unfiltered regions are most likely what observers are matching, and it could be what they use to extract the point and magnitude of the convergence induced by a transparent object. However, other sources of information could be at play, which might correlate with or trade off with the RMC, and more work will be needed to determine their effects, as well as any Gestalt principles at work and how eye movements are used to sample the relevant information. At the least, though, our results support the conclusion that the RMC has a substantial effect on the color of a transparent object.