Scene statistics in 3D natural environments
MetadataShow full item record
In this dissertation, we conducted a stereoscopic eye tracking experiment using naturalistic stereo images. We analyzed low level 2D and 3D scene features at binocular fixations and randomly selected places. The results reveal that humans tend to fixate on regions with higher luminance variations, but lower disparity variations. Because of the often observed co-occurrence of luminance and depth changes in natural environments, the dichotomy between luminance features and disparity features inspired us to study the accurate statistics of 2D and 3D scene properties. Using a range map database, we studied the distribution of disparity in natural scenes. The natural disparity distribution has a high peak at zero, and heavier tails that are similar to a Laplace distribution. The relevance of natural disparity distribution to other studies in neurobiology and visual psychophysics are discussed in detail. We also studied luminance, range and disparity statistics in natural scenes using a co-registered luminance-range database. The distributions of bandpass 2D and 3D scene features can be well modeled by generalized Gaussian models. There are positive correlations between bandpass luminance and depth, which can be captured by varying shape parameters in the probability density functions of the generalized Gaussians. In another study on suprathreshold luminance and depth discontinuities, we show that observing a significant luminance edge at a significant depth edge is much more likely than at homogeneous depth surfaces. It is also true that a significant depth edge happens at a significant luminance edge with a greater probability than at homogeneous luminance regions. Again, the dependency between luminance and depth discontinuities can be modeled successfully by generalized Gaussians. We applied our statistical models in 3D natural scenes to stereo correspondence. A Bayesian framework is proposed to incorporate the bandpass disparity prior, and the luminance-disparity dependency in the likelihood function. We compared our algorithm with a classical simulated annealing method based on heuristically defined energy functions. The computed disparity maps show great improvements both perceptually and objectively.