RI: Small: Performance Prediction and Validation for Object Recognition

该项目汇集了工程学和统计学之间的跨学科合作，以发展自动物体识别科学。物体识别技术在安全和执法、国防和安保、自主导航、工业制造、业务流程和电子商务等广泛应用中无处不在。提出的变革性研究为基于模型和数据的概率和结构特征预测目标识别算法的性能提供了一个基础框架。与以往的研究相比，本文不仅考虑了数据失真因素，而且考虑了模型相似度。性能被明确地建模为数据失真因素（特征不确定性、遮挡和杂波）和模型因素（相似性）的函数，因此人们可以最终表征性能的概率分布，而不是特定数据集上的经验结果。研究将集中在三个方面：1)发展性能预测和性能边界的贝叶斯公式。2)分析绩效预测方法中所有假设的效果；评估数学可追溯性、复杂性和性能增益的影响，并根据实际数据验证不同的假设。3)为实际应用生成该方法的结果。识别的科学理论、预测和计算模型的发展将导致识别系统设计的系统方法的发展，这些方法可以在复杂的现实世界中可靠地实现可预测的结果，并有助于基于计算机的物体识别科学。

This project brings together an interdisciplinary collaboration between Engineering and Statistics for developing the science for automated object recognition. Object recognition technology is pervasive in a broad range of applications, such as safety and law enforcement, defense and security, autonomous navigation, industrial manufacturing, business process and e-commerce. The proposed transformative research provides a foundational framework for predicting the performance of object recognition algorithms based on probabilities and structural characteristics of the models and data. In contrast to previous work, the proposed research considers not only the data distortion factors but also the model similarity. The performance is explicitly modeled as a function of data distortion factors (feature uncertainty, occlusion and clutter) and model factors (similarity) so that one can ultimately characterize the probability distribution of performance rather than the empirical results on a specific dataset. The research will focus in three areas: 1) Developing the Bayesian formulation of performance prediction and bounds on performance. 2) Analyzing the effects of all of the assumptions made in the performance prediction approach; evaluating the effects of mathematical tractability, complexity and the gain in performance and validating different assumptions based on real data. 3) Generating results of the approach for practical applications. The development of scientific theory, prediction and computational models for recognition will result in the development of systematic approaches to the design of recognition systems that can reliably achieve predictable results in complex real-world, and contribute towards the science of computer-based object recognition.