Segmentation-based reflectance recovery

The reflectance properties of a surface are an essential factor in its appearance. Much previous work has focused on the problem of reflectance recovery from images. These methods must assume an a priori grouping of pixels into uniform-reflectance regions. In this paper we presented a method for automatic grouping of pixels for reflectance estimation. First a over-segmentation is achieved by traditional image segmentation. For each image region of the over-segmentation, a probability distribution is built and a reflectance subspace is formed by likelihood thresholding. The regions with the same reflectance are then merged by adapting a traditional bayesian formulation for image segmentation to increase estimation accuacy. After completing the merging process, reflectance parameter estimates are computed for the remaining subspaces by the maximum likelihood reflectance estimate. The experiment results on a synthetic scene and a real scene show our method can achieve a more accurate image segmentation and reflectance estimation than traditional methods.