Evaluation of color spaces for edge classification in outdoor scenes

Knowing which edges in an image denote shadow edges and which are due to object boundaries or changes in surface reflectance has important applications in both computer vision and mixed reality. We show that the choice of color space has a significant effect on our ability to differentiate shadow edges from reflectance edges, particularly in sunlit scenes. We have evaluated the performance of 11 color spaces on an input data of more than a hundred colors imaged in a variety of illumination conditions. We quantify the performance of these color spaces using receiver operating characteristic curves, and use the z statistic to find if the difference in performances is statistically significant.