Abstract:  The problem of weakly supervised human body detection under difficult poses (e.g., multi-view and/or  arbitrary poses) is studied. Most current methods on human body detection focuse only on a few common human body poses with human body in upright positions, while in the real world human bodies may exhibit very rich pose variations (e.g., when people are bending, sleeping or sitting). This not only imposes great challenges on the task of human detection, but also makes the job of manual annotation even more difficult, and usually only weak annotations are available in practice. The multiple instance learning method relaxes the requirements of accurate labeling and hence is commonly used to address the task. However, it is sensitive to the quality of positive instances and the settings of some model parameters such as the strategy to fuse the instance-level conditional probability into a bag-level one. This paper presents a comprehensive and in-depth empirical method of these important but less studied   issues on the person dataset of Pascal VOC 2007, and proposes a new selective weakly supervised detection algorithm (SWSD). Experiments demonstrate that with only a few fully supervised samples, the performance of weakly supervised human body detection can be significantly improved under the multiple instance learning framework.

Key words: weakly supervision, human body detection, arbitrary poses, multiple instance learning

摘要： 困难姿态（多视角或者任意姿态）下的弱监督人体检测问题被关注研究。现在大部分人体检测仅仅关注普通的直立姿态，但现实中的人体却呈现非常丰富的姿态（如弯曲的、躺着的、坐着的），这不仅加大了人体检测的难度，而且令标注工作更加困难，实际中通常只能获得弱标注样本。多示例学习方法放松了精准标注的要求，因此常常被用来解决此类问题。但是多示例学习对正示例的质量以及一些模型参数设置相当敏感，例如将示例层次条件概率融合到包层次的策略。在Pascal VOC 2007的人类数据集上对这些重要但很少被关注的问题进行了综合性深度研究，并提出了一种新的选择性弱监督检测算法（selective weakly supervised detection，SWSD）。实验证明，只要添加少量的监督样本，在多示例学习框架下，可以大幅度提高弱监督人体检测性能。

关键词: 弱监督, 人体检测, 任意姿态, 多示例学习