The contribution of color to detect edges in natural scenes
53.437, Tuesday, May 14, 8:30 am - 12:30 pm, Orchid Ballroom
Thorsten Hansen1, Karl Gegenfurtner1; 1General Psychology, Justus Liebig University Giessen
Introduction In a statistical analysis of over 700 natural scenes we found that chromatic edge contrast is statistically independent of achromatic edge contrast and thus is an independent source of information that can be linearly combined with other cues for the proper segmentation of objects (Hansen and Gegenfurtner, 2009, Visual Neuroscience, 26, 3549). Here we investigated to what degree humans use this information. Methods We used four freely available data sets of human marked edges (ANID, BSD100, BSD300, SOD). We converted the images to DKL to separate chromatic from achromatic information in a physiologically meaningful way. Edges were detected in the three planes of the DKL color space and compared to human-labeled edges using ROC analysis for a threshold-independent evaluation. Performance was quantified by the difference of the area under the ROC curves. Results were highly consistent across all data sets. The average improvement using chromatic edges in addition to achromatic edges was about 3% on average but reached up to 11% for some images. Performance dropped to about 2% if only a single additional chromatic plane was used. This small benefit matches the performance of dichromats in psychophysical experiments. Interestingly, almost the same benefit of chromatic information (2.5%) occurred for human-marked edges in gray-scale images. Observers probably use high-level knowledge to correctly mark edges even in the absence of achromatic contrast. Summary The advantage of the additional chromatic channels was small on average (about 3%) but reached up to 11% for some images. Overall, color was advantageous for about 90% out of the images. Conclusion We interpret our results such that edge detection benefits on average from chromatic information, and that this benefit can be very high in some cases.