ANR DEEPVISION: Seeing and understanding humans with deep structured models

计算机视觉的进步使几年前似乎不可能实现的应用成为可能，例如实时手势识别、移动设备上的人脸检测和环境的3D映射。机器学习是这一发展背后的主要推动力。与视觉信息相对应的海量数据，以及由于不同的视角、形状等而导致的固有的巨大变异性，使得学习成为一种吸引人的方法。 然而，目前的技术水平受到限制。结构关系使许多现实情况超出了商业应用的范围，例如人与人和人与物的相互作用，长期的动态行为，以及对高度变形和关节物体的识别。 我们建议通过创建深层结构化模型来应对这些挑战，将深度学习的优势(一系列强大的模型和算法)与结构化模型相结合，后者非常适合对复杂数据进行建模。

Advances in computer vision have enabled applications that seemed impossible some years ago, e.g. gesture recognition in real time, face detection on mobile devices and 3d mapping of environments. Machine learning is a major driving force behind this development. The vast amount of data corresponding to visual information, as well as its inherently large variability due to different viewpoints, shapes, etc. make learning an appealing approach. However, the current state of the art suffers from limitations. Structural relationships put many realistic situations out of reach of commercial applications, such as person-person and person-object interactions, long-term dynamical behaviour, and the recognition of highly deformable and articulated objects. We propose to address these challenges by creating deep structured models, combining the advantages of deep learning, a powerful family of models and algorithms, with structured models, which are well-suited for modeling complex data.