CAREER: Representing, Understanding, and Enhancing Scenes at the Internet-Scale

摄影对社会有着巨大的影响——它是我们主要的视觉历史，也是讲故事、娱乐和艺术的媒介。但是我们的视觉世界非常复杂，这使得计算机视觉很难理解照片，计算机图形学很难合成视觉内容。然而，近年来互联网规模的图片集的出现为研究提供了新的方向。我们使用基于场景的表示来利用互联网规模的数据。场景（地点或环境）是所有其他视觉现象存在的背景，似乎有可能对场景空间进行暴力处理——在数百万个场景中，我们找到了定性相似的场景，并创建了大量数据驱动的算法，其功能与典型的自下而上的图形和视觉管道互补。本研究的基本原则是，场景表示和大型图像数据库的联合研究将推动图形和视觉领域的最新进展。首先，我们正在研究细节合成任务，以减轻相机抖动，运动模糊，散焦，大气散射，或低分辨率。场景表示足够健壮，即使在存在戏剧性模糊的情况下，也能在互联网规模的照片集中找到匹配的场景。这些匹配场景提供了一个特定于上下文的统计模型，该模型可用于插入令人信服的纹理和物体细节。其次，我们正在研究基于属性的场景表示。我们使用众包来发现属性并为社区建立大型数据库。属性是下一代大数据成像研究的一个强大的中间表示，它可以通过机器人、安全、视力受损援助和车辆安全等应用产生广泛的社会影响。研究人员还在为布朗大学的学生开发一门新的入门课程，以探索跨科学学科的大数据计算，并正在创建一个视觉计算教育的在线社区，以使对摄影和编程感兴趣的学生受益。

CAREER: Understanding, Representing, and Enhancing Scenes at the Internet-scalePhotography has an enormous impact on society -- it is our primary visual history and a medium for storytelling, entertainment, and art. But our visual world is extraordinarily complex which makes it difficult for computer vision to understand photos and for computer graphics to synthesize visual content. However, the emergence of Internet-scale photo collections in recent years enables new research directions. We use scene-based representations to leverage Internet-scale data. Scenes (places or environments) are the context in which all other visual phenomena exist and it seems possible to brute-force the space of scenes -- with millions of scenes, we find qualitatively similar scenes and create massively data-driven algorithms with capabilities that are complementary to typical bottom-up graphics and vision pipelines. The underlying principle of this study is that joint investigations of scene representations and large image databases will advance the state-of-the-art in graphics and vision. First, we are investigating detail synthesis tasks which alleviate camera shake, motion blur, defocus, atmospheric scattering, or low resolution. Scene representations are robust enough to find matching scenes in Internet-scale photo collections even in the presence of dramatic blurring. These matching scenes provide a context-specific statistical model which can be used to insert convincing texture and object detail. Second, we are studying attribute-based representations of scenes. We use crowdsourcing to discover attributes and build large databases for the community. Attributes are a powerful intermediate representation for the next generation of big data imaging research which can have broad societal impact through applications such as robotics, security, assistance to vision-impaired, and vehicle safety. The investigators also are developing a new introductory course for Brown students to explore big data computing across scientific disciplines and are creating an online community for visual computing education to benefit students interested in photography and programming.