模型假设权重的三维运动分割

3D Motion Segmentation Based on Model Hypothesis Weights

提出一种使用模型假设权重来有效地构建仿射矩阵的方法。首先,对输入视频序列中的每个图像对进行随机采样,生成模型假设;其次,用生成的模型假设构建一个初始仿射矩阵;然后,计算这些模型假设的权重,并通过信息理论来选择有意义的模型假设,以及用它们的权重计算初始仿射矩阵的权重;接着,由初始仿射矩阵和上述权重得到该图像对的仿射矩阵;最后,累加所有图像对的仿射矩阵得到该视频序列的仿射矩阵,在仿射矩阵上用聚类方法把属于不同运动物体的特征点聚成不同的类。实验结果表明,所提出方法能够获得较高的分割准确率。

This paper proposes a method to effectively construct an affine matrix using model hypothesis weights.First,random sampling is performed on each image pair in the input video sequence to generate model hypotheses.Second,an initial affine matrix is constructed using the generated model hypotheses.Third,the weights of these model hypotheses are calculated,and meaningful model hypotheses are selected by information theory,as well as their weights are used to calculate the weights of the initial affine matrix.Fourth,the affine matrix of the image pair is obtained from the initial affine matrix and the above weights.Finally,the affine matrices of all image pairs are accumulated to obtain the affine matrix of the video sequence,and a clustering method are used on the affine matrix to cluster the feature points belonging to different moving objects into different classes.The experimental results show that the proposed method can achieve a high segmentation accuracy.