Structured Models for Human Activity Recognition

近年来，计算机视觉领域取得了长足的进步。 自动解释各种识别问题的准确性和能力-例如对象识别，人类动作识别-已经大大提高。 这一进展是由新的问题定义和表示、强大的机器学习算法的复兴以及对大量标记训练数据的访问所推动的。然而，目前的算法缺乏产生准确的结构化表示，允许详细的场景理解的能力。 例如，关于场景中的人群或视频序列内的时空关系的推理仍然具有挑战性。 我们建议研究通过关注以下方面来推动这一领域的最新技术水平：（1）用于识别的新型深度结构化模型和（2）促进数据收集并将性能推向100%正确的计算机视觉的主动学习算法。

Great strides have been made within computer vision in recent years. Accuracy and capacity for automatic interpretation on a variety of recognition problems -- e.g. object recognition, human action recognition -- have improved dramatically. This progress has been driven by novel problem definitions and representations, the resurgence of powerful machine learning algorithms, and access to large quantities of labeled training data.******However, current algorithms lack the capability to produce accurate structured representations that permit detailed scene understanding. For example, reasoning about groups of people in a scene or spatio-temporal relations within a video sequence remain challenging. We propose research to push the state of the art in this area by focusing on: (1) novel deep structured models for recognition and (2) active learning algorithms for facilitating data collection and pushing performance toward 100% correct computer vision.********