CAREER: Situated Recognition: Learning to understand our local visual environment

这个项目开发计算机视觉技术来识别我们日常生活中的物体。为了识别我们周围的视觉内容，相机可以在一段时间内记录多个图像，这就有机会利用网络图像无法获得的上下文。这个项目追求新的表示和计算策略，有效地利用这个上下文，在我们的环境中实现高质量的视觉识别。与使用这种上下文的机会相平衡的是在任何特定环境中，面对杂乱、遮挡、非规范视图和特殊外观变化，使识别工作的挑战。所开发的方法可以成为开发技术的核心部分，以帮助计算机视觉系统扩展到识别我们日常世界中的一切。这项研究将导致自动化系统更好地理解和监测我们的日常环境，改善人机交互，并鼓励在这一领域进行更多的研究。这个研究方向不同于大多数的识别工作，主要集中在从网络上收集的互联网图像。这种网络收集的偏差可能导致模型不能推广到特定的环境。位置识别允许利用本地环境，包括人类互动和口语，来建立特定于环境的模型，以及环境中对人们重要的部分。该项目收集多个数据集，强调多视图图像和对环境的长期观察，同时采样各种设置。研究团队开发了通过利用上下文、高效重用和上下文依赖显著性来解析和检测对象的算法；并使用情境自然语言来驱动视觉识别模型的自动学习。项目网页：http://acberg.com

This project develops computer vision technologies for recognizing objects in our daily lives. For recognizing visual content around us, where cameras can record multiple images over a period of time, there is an opportunity to take advantage of context that is not available for internet images. This project pursues new representations and computational strategies exploiting this context efficiently to achieve high-quality visual recognition in our environment. Balanced against the opportunity of using this context is the challenge of making recognition work in any particular environment, in the face of clutter, occlusion, non-canonical views, and idiosyncratic appearance variation. The methods developed can be a core part of developing technology to help computer vision systems scale to recognize everything in our daily world. The research leads to automated systems for better understanding and monitoring of our daily environment, improved human-computer interaction, and encourages more research in this area.This research direction is different from the majority of work in recognition that has focused on internet images collected from the web. The biases of such web-collected may lead to models that do not generalize to a particular environment. Situated recognition allows exploiting local context, including human interaction and spoken language, to build models specific to an environment and furthermore to the parts of an environment that are important to people. The project collects multiple datasets stressing multi-view imagery and long-term observation of environments while sampling a wide variety of settings. The research team develops algorithms to parse and detect objects by exploiting context, efficient re-use, and context-dependent saliency; and uses situated natural language to drive automatic learning of visual recognition models.Project Webpage: http://acberg.com