The Warp Factory

1. Principle The programme of projects now under development under the title 'Warp Factory' aim to revisit one of the most fundamental issues in graphics, namely the representation of images. Partners already involved in Warp Factory projects include the Universities of Bath and Sussex and possibly shortly the University of Burgundy (the Le2i Imaging Lab in Dijon). The idea is to re-cast sampled images into a continuous representation which is as close as possible to the original form from which the samples were taken. The outstanding research questions are 1, whether the continuous-space approximations we can derive from the samples are close enough to the original so that re-sampling results in a visually indistinguishable image and 2, whether we can exploit the continuous-space representation in ways not open to the sampled form, including the application of that representation to data derived from images. Sampled formats are invariably used where the manipulation of photographic or video (live-capture) material is involved. While this is a sensible input-output format it can induce artefacts into the results of any attempts at manipulating a sampled image directly (e.g. warps). We claim that a form of boundary representation for continuous spaces is more suitable for the image-manipulation tasks we have identified here and many others, particularly those involving live-capture images. Further, the continuous-space representation is a resolution-independent form, so lends itself to processes which operate on multi-resolution image representations. There are, however, a number of problems that have to be solved along the way. We aim to work on entire sequences of images rather than single images in the belief that we can exploit frame-to-frame continuity. We will be assuming that we have 'clean' and relatively noise-free images, although matching sequences gives us additional protection from non-systematic noise processes. Finally we will be accepting that the costs of semi-automatic process can more easily be amortised (and justified) over sequences rather than over single images. The imaging model we want to use is based on continuous isochromic contours in Lab-space [Wysz&82]. That is, the image is defined in terms of closed contours that specify the exact boundaries for regular intervals of colour in a style analogous to height boundaries in terrain contour maps. For Lab-space that means one contour map for L (greyscale) and one each for a and b (colour parameters along yellow-blue and red-green axes respectively). The contours are defined in terms of non-uniform B-splines …