Abstract

Generative Modeling for Continuous Non-Linearly Embedded Visual Inference
Cristian Sminchisescu - University of Toronto Allan Jepson - University of Toronto
Many difficult visual perception problems, like 3D human motion estimation,can be formulated in terms of inference using complex generative models,defined over high-dimensional state spaces. Despite progress, optimizing suchmodels is difficult, because prior knowledge cannot be flexibly integrated inorder to reshape an initially designed representation space. Non-linearities,inherent sparsity of high-dimensional training sets, and lack of globalcontinuity makes dimensionality reduction challenging and low-dimensionalsearch inefficient. To address these problems, we present a learning andinference algorithm that restricts visual tracking to automatically extracted,non-linearly embedded, low-dimensional spaces. This formulation produces alayered generative model, with reduced state representation, that can beestimated using efficient continuous optimization methods. Our priorflattening method allows a simple analytic treatment of low-dimensionalintrinsic curvature constraints, and allows consistent interpolationoperations. We analyze reduced manifolds for human interaction activities, anddemonstrate that the algorithm learns models that are useful for tracking andfor the reconstruction of 3D human motion in monocular video.

Generative Modeling for Continuous Non-Linearly Embedded Visual Inference

Cristian Sminchisescu - University of Toronto
Allan Jepson - University of Toronto

Many difficult visual perception problems, like 3D human motion estimation,can be formulated in terms of inference using complex generative models,defined over high-dimensional state spaces. Despite progress, optimizing suchmodels is difficult, because prior knowledge cannot be flexibly integrated inorder to reshape an initially designed representation space. Non-linearities,inherent sparsity of high-dimensional training sets, and lack of globalcontinuity makes dimensionality reduction challenging and low-dimensionalsearch inefficient. To address these problems, we present a learning andinference algorithm that restricts visual tracking to automatically extracted,non-linearly embedded, low-dimensional spaces. This formulation produces alayered generative model, with reduced state representation, that can beestimated using efficient continuous optimization methods. Our priorflattening method allows a simple analytic treatment of low-dimensionalintrinsic curvature constraints, and allows consistent interpolationoperations. We analyze reduced manifolds for human interaction activities, anddemonstrate that the algorithm learns models that are useful for tracking andfor the reconstruction of 3D human motion in monocular video.