Incremental Learning of Object Categories

复杂视觉场景中物体的自动识别仍然是计算机视觉领域最活跃的研究领域之一。然而，目前大多数可用的系统只能使用固定数量的已知对象类（目前在100到1000之间）来解决非常具体的问题。该建议通过开发方法来克服静态系统的这些限制，该方法通过增量增加其知识来适应最初从固定数据集训练的系统。我们把这种随时间调整系统知识的过程称为“增量学习”或“终身学习”。这些方法本质上是用于识别和表示以前未见过的类别，特别是在万维网（如Flickr）上公开可用的图像集合的时代。该项目的长期目标是缩小人类和机器视觉之间的差距，这可以通过构建能够自动检测新对象类别并将其添加到当前数据集中的系统来实现。在这个过程中，调整当前的表示方式将是一个重要的主题，以确保不同类别的可分离性，即使有数千个已知的类。在项目的第一阶段，新颖性检测和从少数例子中学习将是研究的重点。此外，需要在特征层和对象实例层研究有效更新当前模型的方法。因此，可以避免代价高昂的从头开始学习。此外，主动学习是一个高度相关的主题，以确保新示例与最终分类模型的相关性。由于在这一研究领域没有通用的基准数据集，我们计划建立一个数据集，并定义终身学习场景的评估标准，这些标准将公开提供。作为现实世界的应用，我们将在移动机器人上评估开发的方法，该机器人将自主探索其环境并逐步扩展其知识。这个建议的长期目标是开发方法，使系统能够自动适应日常变化的世界，并随着时间的推移增加和纠正其知识。

Automatic object recognition in complex visual scenes is still one of the most active research areas in computer vision. However, the majority of currently available systems are only capable of solving very specific problems with a fixed number of known object classes (currently between 100 and 1000). This proposal overcomes these limitations of static systems by developing methods, which adapt a system initially trained from a fixed dataset by incrementally increasing its knowledge. We call this process of adapting a systems knowledge over time "incremental learning" or "lifelong learning". Such methods are essentially for recognizing and representing previously unseen categories, especially in the time of publicly available image collections on the world wide web such as Flickr. A long term goal of this project is to close the gap between human and machine vision, which can be done by building systems that are able to automatically detect new object categories and to add them to their current datasets. Adapting the current representation modalities will be one important topic involved within this process to ensure the separability of different categories even with thousands of known classes. In the first stage of the project, novelty detection and learning from few examples will be in the focus of research. In addition, methods for efficiently updating the current model need to be investigated, both on the level of features and on the level of object instances. Thereby, costly learning from scratch can be avoided. Furthermore, active learning is a topic highly relevant to ensure the relevance of new examples with respect to the resulting classification model. Since there is no common benchmark dataset within this area of research, we plan to build a dataset and to define evaluation criteria for lifelong learning scenarios which will be made publicly available. As an real-world application, we will evaluate the developed methods on a mobile robot, that will autonomously explore its environment and will incrementally extend its knowledge. The long term goal of this proposal is to develop methods, with which a system is able to automatically adapt itself to the daily changing world and to increase and correct its knowledge over time.