Weakly Supervised Learning for Depth Estimation in Monocular Images

单目图像深度估计的弱监督学习

• 批准号: 420493178
• 负责人: Professor Dr. Ralph Ewerth
• 金额: --
• 依托单位: Technische Informationsbibliothek (TIB)
• 依托单位国家: 德国
• 项目类别: Research Grants
• 财政年份: 2019
• 资助国家: 德国
• 起止时间: 2018-12-31 至 2021-12-31
• 项目状态: 已结题
• 来源: https://gepris.dfg.de/gepris/projekt/420493178?language=en
• 关键词: Weakly; Supervised; Learning; Depth; Estimation

Leveraging two types of machine learning methods, namely learning to rank and superset learning, this project develops novel approaches to depth estimation in monocular images. Both approaches merely require "weak" supervision, either in the form of relative ("object B is behind object A") or rough absolute depth information ("object A is close to the camera"), and thereby facilitate the acquisition of training data. As predictions, they produce qualitative depth maps in the form of rankings, specifying the relative order of objects in a scene. This is in contrast to conventional approaches based on statistical regression, which require precise training data and produce (unnecessarily) precise predictions. For both approaches, machine learning algorithms specifically tailored for the problem of depth estimation will be developed. These will be combined with two approaches to feature construction: the systematic (hand-crafted) modeling of monocular depth clues of human perception, and the use of deep neural networks for representation learning. Our qualitative, weakly supervised approaches to monocular depth estimation will be analyzed and compared with each other, as well as with existing approaches based on statistical regression. Last but not least, the benefits of our new algorithms will be investigated for several important applications, namely visual concept classification in images and videos, visual concept detection (by means of localization), and image segmentation.

利用两种类型的机器学习方法，即学习排名和超集学习，该项目开发了单目图像深度估计的新方法。这两种方法都只需要“弱”监督，以相对（“对象B在对象A后面”）或粗略的绝对深度信息（“对象A靠近相机”）的形式，从而便于获取训练数据。作为预测，它们以排名的形式产生定性的深度图，指定场景中对象的相对顺序。这与基于统计回归的传统方法形成对比，后者需要精确的训练数据并产生（不必要的）精确预测。对于这两种方法，将开发专门针对深度估计问题定制的机器学习算法。这些将与两种特征构建方法相结合：人类感知的单眼深度线索的系统（手工）建模，以及使用深度神经网络进行表示学习。我们的定性，弱监督的单目深度估计方法将进行分析和相互比较，以及与现有的基于统计回归的方法。最后但并非最不重要的是，我们的新算法的好处将被调查的几个重要的应用，即视觉概念分类的图像和视频，视觉概念检测（通过本地化），图像分割。