Embedding view-dependent covariance matrix in object manifold for robust recognition
Embedding view-dependent covariance matrix in object manifold for robust recognition
复制标题
在对象流形中嵌入依赖于视图的协方差矩阵以实现鲁棒识别
DOI：
--
复制
发表时间：
2008
期刊：
影响因子：
--
通讯作者：
H. Murase
中科院分区：
文献类型：
--
作者：
Lina Tomokazu Takahashi;I. Ide;H. Murase
相机捕捉到的图像的变化通常是自然发生的。例如，对于每个姿势，对象的外观通常不同，并且在捕捉过程中可能会出现降级效果。虽然我们可以用一个简单的流形来表示姿态的可变性，但依靠简单的流形技术来同时处理姿态和退化问题是不可能的，因为简单的流形没有考虑样本在特征空间中的分布信息。本文提出了一种在目标流形中嵌入视点相关协方差矩阵的方法来开发一个稳健的三维目标识别系统。这里，通过沿流形内插特征向量和特征值，以一种有效的方式获得视点相关协方差矩阵。实验结果表明，我们开发的三维物体识别系统即使在受到几何失真和质量退化影响的图像中也能准确地识别三维物体。
Variations in camera-captured images usually occur naturally. For example, the appearance of an object usually differs for every pose and degradation effect might occur during the capturing process. While we could use a simple manifold to represent the variability of pose, relying on the simple manifold technique to deal with both pose and degradation problems is not possible, since a simple manifold does not take into account the information of sample distributions in feature space. In this paper, we propose a technique which embeds viewdependent covariance matrix in object manifold to develop a robust 3D object recognition system. Here, the view-dependent covariance matrices were obtained in an efficient way by interpolating eigenvectors and eigenvalues along the manifold. Experiment results showed that our developed 3D object recognition system could accurately recognize 3D objects even from images which are influenced by geometric distortions and quality degradation effects.