Digital video and film manipulation has been the cornerstone of video broadcast and cinema industry for some time now. However transparency in video remains a challenge that prevents video manipulation techniques from being used ubiquitously for data editing. For instance in the vast literature on frame rate conversion motion compensated interpolation is a key component. When faced with transparency however, new frames cannot be built reliably using the usual assumption of one moving object per pel. Archive restoration is another example where transparency is an issue. Traditional techniques assume the examined corruption is opaque, completely obliterating the underlying original data. However corruption is in fact transparent. Ignoring this observation often generates artifacts during restoration.
This PhD shows that video processing techniques can be improved by addressing transparency in an explicit form. We focus on three main problems: 1) Archived Footage Restoration, 2) Transparency Detection and 3) Motion Estimation for Regions of Transparency. A summary of the findings is available here. Full thesis can be accessed here .
