CAREER: Learning to See - A Unified Segmentation and Recognition Approach

图像分割、形状检测和目标识别是三个紧密耦合、相互交织的视觉过程。只有将它们集成到一个连贯、统一的系统中，才有希望实现大规模视觉识别的长期目标。该建议提出了一种基于学习的统一图形表示法，为解决这一极其困难的问题提供了一个原则性和可操作性的框架。本文的研究工作主要集中在三个方面：1)通过构造用于形状检测、目标识别和分割的多层图来完成目标信息的推理；2)自下而上的底层线索与自上而下的目标形状知识的无偏集成；3)利用有监督的谱图切割学习技术直接学习目标识别图。拟议的100个和1000个对象识别挑战赛将对其成功进行高度客观和严格的评估。这项研究借鉴了不同学科的想法：计算机视觉、机器学习、数值分析和计算理论。PI将继续通过研究生课程、会议教程和讲习班、本科生实习和研究经验以及提供教程和开放源代码的网页资源来促进跨学科研究。

Image segmentation, shape detection and object recognition are three tightly coupled and intertwined visual processes. Only by integrating them together into a coherent, unified system can there be hope for achieving the long-standing goal of large-scale visual recognition. This proposal puts forward a learning-based unified graph formulation providing a principled and workable framework for attacking this extremely difficult problem. The proposed research efforts are directed along three fronts: 1) complete inference of object information by constructing multi-layer graph for shape detection, object recognition and segmentation; 2) unbiased integration of bottom-up low level cues with top-down knowledge of object shape; and 3) direct learning of the object recognition graph with supervised spectral graph cuts learning technique. The proposed 100- and 1000- Object Recognition Challenge will provide a highly objective and rigorous evaluation its success. This research draws upon ideas from a diverse set of disciplines: computer vision, machine learning, numerical analysis, and theory of computation. PI will continue to promote inter-disciplinary researches through graduate courses, conference tutorials and workshops, internships and research experience for undergraduates, as well as web page resources offering tutorial and open source code.