Feature matching is a vital stage in many image processing applications. Finding accurate correspondences is made difficult by phenomena such as occlusions, non-rigid deformations, motion blur and more. We posit that some scenarios do not have enough information for an accurate automatic solution. Although many applications are required to be automatic, there are others that can benefit from being semi-automatic, allowing the user to provide assistance to areas where the system is failing. Good

more »

... examples of this exist in the media post-production world, such as multi-view scene reconstruction, sparse-to-dense disparity estimation from view matching, image mosaic'ing (digital panoramas), or even motion estimation. The presented paper describes how to incorporate user-assistance into a Bayesian feature matching framework. By adding user information in the form of intuitive Bezier curves, difficult regions can be matched with the same accuracy as easier to match areas. The presented system uses a simple optimisation scheme, giving the user real-time interactive control over the corrected matches.