3d Video Format Conversion

The main aim of this thesis work is to find/implement various methods that convert Conventional Stereoscopic 3D Video (CSV) to Multiview video (MVV). The work investigates different methods that can produce multiple views given a stereoscopic pair from a frame of a particular video sequence and continues with the process of selecting the best among investigated methods that has optimum quality and speed. In contrast to the existing algorithms, this work disregards the physical depth but instead focus on pixel value correspondence. The intermediate view generation in this work is not considered as a geometrical problem, but a morphing problem. Different morphing algorithms (mesh, field and thin plate spline morphing techniques) are considered for conversion. Performance of each morphing algorithm is in turn compared using different correspondence matching techniqes. The investigated methods aim to produce arbitrary number of novel synthesized camera views from a sparse view set. Mesh morphing algorithm is found to be a better candidate in terms of signal to noise ratio, but requires accurate correspondences at edges of an object in a particular scene and also needs more execution time to generate more number of views. A new approach to field morphing has been introduced in this thesis work, which performs better in terms of execution time and also found to produce intermediate views with reasonable signal to noise ratio. This approach is observed to bring good trade off between speed and accuracy. This conversion has an advantage it can be used as a decompression mechanism that can produce multiple views required for an Autostereoscopic 3D display from a stereoscopic left and right pair. This approach also brings the benefit of backward compatibility as present standards for CSV may be used to provide multiview 3D video to high fidelity Autostereoscopic 3D displays of the future. This work has applications in free view point television, video conferencing systems etc.,