Multiresolution three-dimensional range segmentation using focus cues

This dissertation deals with a problem of reconstructing the 3-D structure of a scene given only 2-D images. We define the problem as a range segmentation problem using focus cues. The objective of the range segmentation is to partition a scene into 3-D regions with different depth ranges to produce a description of the 3-D structure of the scene, without initial depth estimates of objects. We use a finite number of images with different focus positions. Focus cues are used to measure depth ranges, and segmentation is performed using a multiresolution approach.;We first perform a range classification stage to obtain quantized range estimates. The range classification is performed by combining two paradigms: focus cues and Bayesian estimation. A criterion function computed from focus information provides a basic method for measuring the distribution of ranges in each region of a scene. A Bayesian estimate is obtained by modeling the class field as a Markov random field (MRF). An energy functional of the Gibbs distribution of the class field is defined by the MRF-Gibbs equivalence. To combine these two paradigms, we define a combined energy functional in terms of the energy functional from the criterion function values for focus measure and the energy functional of the Gibbs distribution of the class field. Then the combined energy functional is minimized by a modified simulated annealing method to obtain range classification. The range classification gives quantized range estimates, and it also provides an initial range segmentation.;For range segmentation, we obtain interpolated range values and 3-D vectors in a world coordinate system. Range segmentation is composed of two phases: a merging process and a multiresolution range segmentation. First, a merging process is performed to merge initial segments from the range classification result. This gives a three-dimensional range segmentation at the coarsest resolution. Second, three-dimensional multiresolution range segmentation (3D MRS) is performed to refine the range segmentation into finer resolutions. To create a multiresolution framework, energy functionals are defined over multiple resolutions using multiresolution Markov random fields. The 3D MRS algorithm first performs range segmentation at the coarsest resolution and proceeds progressively to finer resolutions.;The proposed range segmentation method does not require initial depth estimates. The range segmentation provides a rich description of the 3-D structure of a scene.