A perceptual organization approach for figure completion, binocular and multiple-view stereo and machine learning using tensor voting.

摘要

This works extends the tensor voting framework and addresses a wide range of problems from a perceptual organization perspective. The most important contributions are the addition of boundary inference capabilities, a novel re-formulation of the framework applicable to high-dimensional spaces and the development of algorithms for computer vision and machine learning problems. In all cases, the problem is formulated as the organization of the inputs into salient perceptual structures.;For single image analysis, we address the inference of integrated descriptions in terms of edges and keypoints in way that can be useful for higher level processes. We also address a higher level problem: figure completion. We propose a computational framework which implements both modal and amodal completion and provides a fully automatic decision making mechanism for selecting between them. We illustrate the approach on several inputs producing interpretations consistent with those of human observers.;We propose an approach for stereovision that considers both binocular and monocular cues. It allows the integration of matching candidates generated by different operators, combining their strengths. Then, perceptual organization is performed in 3-D under the assumption that correct matches, unlike wrong ones, form salient coherent surfaces. Disparity hypotheses for unmatched pixels are generated considering both geometric and photometric criteria. We also address dense, multiple-view stereo under the same assumption. Unlike other approaches, ours can process all data simultaneously, can be applied to more general camera configurations and does not require foreground/background segmentation. We were able to reconstruct scenes that provide serious challenges to state-of-the-art methods.;Finally, we present a new implementation of tensor voting that can be generalized to spaces with hundreds of dimensions, since it achieves significant reductions in storage and computational requirements. Its advantages include its applicability to a far wider range of datasets, noise robustness, absence of global computations and capability to process very large numbers of points. We present results in dimensionality and manifold orientation estimation, geodesic distance measurement, nonlinear interpolation and function approximation. This work opens the door for applications such as unsupervised classification, forward and inverse kinematics.