Abstract:
This thesis documents a work in which prediction of depth related discomfort levels, while watching stereoscopic 3D videos is studied. In commercial 3D videos, excessive depth levels can cause discomfort to the viewer and hence decreases the users quality experience of the video. Therefore detecting excessive levels of depth is important to maintain a better visual quality in 3D videos. In this work a scheme is presented for detection of depth discomforts, resulting from excessive depth levels. An exaggerated depth corresponds to high disparity levels between stereo image pairs. In order to detect depth discomforts, we developed and tested algorithms to detect and track maximum disparities. The maximum disparities are extracted from sparse disparity maps, where the disparities are obtained for only certain edge locations. The sparse disparity maps are obtained using ve varieties of block matching, which are: Sum of Absolute Di erences (SAD), Herman Weyl's Discrepancy Measure (HWDM), Adaptive Support Windows (ASW), Sum of Absolute Di erences of Scale Invariant Feature Transforms (SADSIFT) and Correlation of Gradient Orientations (CGO). A comparative study of these ve methods is performed in terms of their performances in estimation of maximum disparities. Also subjective tests are run by collecting viewer discomfort data and using maximum disparity statistics as a predictor of user experience. By examining our results, we observed CGO performs better in maximum disparity estimation. Also it is shown, that the maximum disparity statistics obtained through CGO can be used to predict the number of viewers, which experience depth discomforts.
