Learning Rich Statistical Models of the Visual World for Robust Perception

强大的智能依赖于对缺失、不完整、模糊和损坏的数据进行推理的能力。这在视觉感知中尤其如此，在视觉感知中，智能系统只需要给出二维图像，就能对不断变化的三维世界的复杂性进行推理。贝叶斯推理已经成为处理这类问题的流行方法，因为它提供了一种将模糊传感器测量与关于世界的先验知识相结合的合理方法。先验代表了感知系统收集到的经验，并且通过以统计上合理的方式整合异质信息源，使这样的系统能够对新情况做出稳健的反应。马尔可夫随机场（mrf）为表示视觉先验提供了一种强大而流行的形式。然而，他们通常只建模局部的，两两像素的相互作用，这限制了他们的建模能力。该项目旨在使用更大的像素邻域（团）来提高这些模型的能力和适用性。所提出的专家领域（field -of- experts, FoE）模型推广了许多先前的MRF模型，并且它的所有参数都可以从真实世界的训练数据中学习到。例如，初步实验表明，图像重建应用受益于这种更丰富的视觉先验，但许多其他应用领域仍未探索。这些统计建模工具的发展也将对机器视觉以外的其他领域产生影响，这些领域需要对复杂的高维数据进行建模。最后，收集到的实验数据、学习模型和软件的传播有望刺激研究，并使对视觉世界更好的统计模型的定量比较成为可能。

Robust intelligence rests on the ability to reason about missing, incomplete, ambiguous and corrupted data. This is particularly true in visual perception, where an intelligent system is faced with reasoning about the complexity of a changing three-dimensional world given only two-dimensional images. Bayesian inference has become popular for dealing with such problems because it provides a sound way of combining ambiguous sensor measurements with prior knowledge about the world. Priors represent the collected experience of a perceptual system and by integrating heterogeneous sources of information in a statistically sound way enable such a system to respond robustly to novel situations.Markov random fields (MRFs) provide a powerful and popular formalism for representing visual priors. However, they have typically modeled only local, pairwise pixel interactions, which limit their modeling capabilities. This project aims at increasing the power and applicability of these models using larger pixel neighborhoods (cliques). The proposed Fields-of-Experts (FoE) model generalizes many previous MRF models, and all its parameters can be learned from real-world training data. Preliminary experiments have shown that, for example, image reconstruction applications benefit from such richer visual priors, but many other application domains have remained unexplored. The development of these statistical modeling tools will also have an impact on other domains outside of machine vision where the need for modeling complex, high-dimensional data arises. Finally, the dissemination of the collected experimental data, learned models, and software promises to stimulate research and make possible quantitative comparisons towards better statistical models of the visual world.