Experimental Performance Characterization of Edge Detection in Computer Vision

随着新的边缘检测论文的出现，令人信服的证据表明，相对于过去的技术，真正的优势对于社区的认可和接受至关重要。这需要建立方法来客观和定量地证明新算法是否提供了性能改进。这项研究将通过自适应地采样边缘检测器参数空间和生成ROC曲线来产生一个精心开发和记录的边缘检测器评估实验框架。这项研究包括记录像素级评估的结果与基于更高级别任务的评估的一致性，例如感知分组、运动结构和人体对象识别。这项工作将留下其他人使用和建立研究所必需的文物(图像集、软件等)。特别是，将创建一个网站，促进将这一框架作为一项标准技术的使用。除了为边缘检测器的性能评估提供解决方案外，这些结果还可以作为开发用于其他计算机视觉问题的性能评估方法的模型。

As new edge detection papers appear, convincing proof of real advantages over past techniques is critical to community recognition and acceptance. This requires establishment of methods to objectively and quantitatively demonstrate whether a new algorithm offers performance improvements. This research will result in a carefully developed and documented experimental framework for edge detector evaluation by adaptively sampling edge detector parameter space and generating ROC curves. The research includes documenting how well the results of the pixel-level evaluation agree with evaluations based on higher-level tasks such as perceptual grouping, structure from motion and human object recognition. The work will leave behind the artifacts (image sets, software, ) necessary for others to use, and to build upon, the research. In particular, a web site will be created that promotes the use of this framework as a standard technique. Beyond providing a solution for performance evaluation of edge detectors, the results may also serve as a model for the development of performance evaluation methods for other computer vision problems.