We present a method for refining disparity maps generated by a stereo matcher for the purpose of computational photography. We automatically fill holes in disparity maps, remove noisy artefacts, and enhance visible object geometries based on available disparity and image data, for the purpose of generating visually appealing depth representations. The key idea is that we use image features (e.g. edges) of the base image (say, the left image of the stereo pair) for enhancing the corresponding depth map. To achieve this, we analyse the base image by spectral matting, and then revise disparity values by a weighted median filter. Experiments show that our method is able to fill holes (i.e. pixels where depth information is unavailable), to revise inaccurate object edges, and to remove speckle noise and invalid step-edges from the given depth information. Besides photo editing, results provided by our method can also be used for image segmentation or object detection.