Abstract

Image analysis and enhancement tools such as tone mapping, colorization, stereo depth, and photomontage, require computing a solution (e.g., for exposure, chromaticity, disparity, labels) over the pixel grid. Computational and/or memory costs often require that a smaller solution be run over a down-sampled image. Although traditional upsampling methods can be used to interpolate the low resolution solution to the full resolution, these methods generally assume a smoothness prior for the interpolation.

We demonstrate that in the cases, such as those above, where a high resolution image is available, it can be leveraged as a prior in the context of a joint bilateral upsampling procedure to produce a better high resolution solution. We show results for each of the applications above and compare them to traditional upsampling methods.