Previous studies in reflectance estimation generally require prior segmentation of an image into regions of uniform reflectance. Due to the measurement noise and limited sampling of the BRDF (bi-directional reflectance function) directions, such estimated results of reflectance are not accurate. In this paper, we propose a novel method for reducing uncertainty in reflectance estimates by merging image regions which have consistent reflectance observations. Each image region acts as a reflectance subspace, so merging of the image regions can result in subspace reduction. We propose a Bayesian segmentation framework to decrease the reflectance uncertainty by using novel merging criteria. Finally, a maximum likelihood reflectance estimation is made for each resulting image region. Experimental results verify the feasibility and superiority of this reflectance-oriented region merging method.