Color Constancy with Faces
This paper presents an empirical study on the problem of solving color constancy using faces. We investigate the effectiveness of using the human face in a scene as the source of chromatic information to achieve color constancy. State-of-the-art algorithms are evaluated to see whether the chromatic information provided by faces may help to estimate the scene illuminant. We create a dataset containing pairs of images of people; each pair includes one image with a neutral-color reference card and the other image without the card. The neutral-color reference card can be used to retrieve the `ground-truth' illuminant of the scene. The dataset covers diverse scenes with various lighting conditions, and the face regions of the people in the images are manually labeled. Based on the dataset, we show that the color on the face highly correlates with the scene illuminant. The results of empirical evaluation on the dataset suggest that, when face information is available, better performance of color constancy can be achieved by inferring the scene illuminant from the face instead of the whole scene.