Object Recognition in Cluttered Scenes Using Appearance-Based Parts and Relationships

点击翻译按钮获取中文摘要

The recognition of general three-dimensional objects in cluttered scenes remains a challenging problem. In particular, the design of a good representation suitable to model large numbers of generic objects that is also robust to occlusion and segmentation problems, has been an stumbling block in achieving success. In this research, a new representation using appearance-based parts (ABPs) and relationships (ABRs) is proposed to overcome these problems. ABPs and ABRs are defined in terms of closed regions, segmented using the Minimum Description Length principle, whose appearance is obtained from collections of images and compactly stored in a hierarchical structure of parametric eigenspaces. This new representation allows the representation of free form objects and has several appealing features, namely (1) robustness to segmentation problems, since it is learned from segmented images; (2) robustness to occlusion steming from the fact that it is based on parts rather than on global properties; and (3) ability to handle large object databases, due to its hierarchical nature. As part of this research a probabilistic model of the discriminatory power of the proposed representation will be also developed. This model will be used in a Bayesian framework, based on the PI's previous work, to design a recognition system. The new representation and the associated probabilistic model, coupled with a Bayesian reasoning engine, will enable the system to automatically recognize and locate generic objects in cluttered scenes. The main contributions of this research are: (1) a new object representation capable of representing large databases of generic objects that is also robust to segmentation problems and occlusion; (2) a probabilistic model of the discriminatory power of the proposed representation and a recognition system based on a Bayesian framework that does not require the use o f ad hoc heuristics; and (3) a rigorous experimental protocol to characterize the performance of the system in the presence of occlusion.