新型确定性模型拟合方法研究及其应用

Model fitting is an important research area in computer vision. The goal of model fitting is to estimate the model instances that effectively explain the information contained in observed data. Model fitting is a challenging problem because that observed data may contain large numbers of outliers, and may also contain more than one model instance. Most mode fitting methods are based on randomized nature, that is, they have no guaranteed stability. For deterministic fitting methods that can provide stable fitting results, this project addresses issues how to improve the algorithmic efficiency and the fitting accuracy. In addition, this project also does the research about the application of model fitting. Summarily speaking, the detailed content of this project is as follows: 1) A novel and fast deterministic fitting method is proposed by analyzing the prior information of co-segmentation; 2) A novel and highly accurate deterministic fitting method is proposed by combining subset sampling and model selection; 3) A novel deterministic fitting based motion segmentation method is proposed. This project is proposed based on the applicant's research foundation on model fitting during his PHD and current research work on computer vision. The project has good feasibility and high academic and applicable values.

模型拟合是计算机视觉中一个重要的研究领域。模型拟合的主要任务是有效地拟合观测数据中所蕴含的所有模型实例。由于观测数据中往往含有大量的离群点且可能同时存在多个模型实例，其研究具有很大的挑战性。现有的模型拟合方法一般是基于随机性的，无法保证拟合结果的稳定性。本项目针对能够提供稳定性拟合结果的确定性拟合方法，研究如何提高该新型拟合方法的计算效率和拟合精度。此外，本项目还研究如何将拟合方法推广到实际应用中。概况地说，本项目的具体研究内容包括：1)通过分析协同分割的先验信息，提出一种快速的新型确定性拟合方法；2)通过融合数据抽样与模型选择，提出一种高精度的新型确定性拟合方法；3)基于确定性拟合，提出一种新型运动分割方法。本项目是申请人在博士期间模型拟合方面的研究基础上，结合现在所在实验室的基础提出来的研究课题,其研究具有广泛的实际意义和学术价值。

模型拟合是计算机视觉中一个重要的研究领域。模型拟合的主要任务是能够有效地拟合观测数据中所蕴含的所有模型实例。由于观测数据中往往含有大量的离群点且可能同时存在多个模型实例，其研究具有很大的挑战性。现有的模型拟合方法在准确性和稳定性还存在较大的问题，本项目针对数据子集采样、模型选择策略和应用推广等问题展开研究。概况地说，本项目的成果如下：1) 通过超图建模刻画模型假设与数据点之间的复杂关系，以提出一种基于模式搜索的模型拟合方法；2) 通过融合数据子集采样和模型选择以提升采样的质量和拟合效果；3) 通过超像素刻画数据点的表观特征，以确定性地获取拟合结果；4) 通过确定性拟合方法挖掘运动轨迹的关联，以有效地获取运动分割结果。提出的模型拟合方法有效地提升了当前的拟合结果，并将模型拟合方法推广到运动分割应用中。发表IEEE TPAMI/TIP，IJCV，PR等权威期刊13篇，1篇CCF A类会议AAAI，授权发明专利3项。