Attribute-based knowledge transfer learning for human pose estimation

作者：Li, Feng; Zhou, Shu-Ren^*; Zhang, Jian-Ming; Zhang, Deng-Yong; Xiang, Ling-Yun

来源：Neurocomputing, 2013, 116: 301-310.

DOI：10.1016/j.neucom.2012.05.032

• Computer vision
• Human pose estimation
• Knowledge transfer learning
• Attribute model
• Supervised Latent Dirichlet Allocation
• Bag of features

Because of the diversity of human poses and appearances of the same pose, it is difficult to collect comprehensive training samples of all kinds of human poses and maintain the probability distribution of the training samples with the same as that of testing. Therefore, this paper learns prior knowledge from the known human pose in the collected samples to infer characteristics of new human poses. This way realizing zero-shot and one-shot learning overcomes the above two difficulties. Finally, a novel human pose estimation based on knowledge transfer learning is proposed in this paper. In the process of our human pose estimation method, first, an attribute-based representation model of the human pose is built based on our proposed "body-pose-attribute" hierarchical framework. Under this model, an image would be divided into disjoint regions, and an attribute is extracted from each region. Thus, an attribute bags can be used to represent a specific human pose, then the attribute bag of a new human pose can be effectively transferred from the prior knowledge obtained from the known human poses. Second, an attribute parameter model called supervised LDA (SLDA) is built, and the Gibbs sampling algorithm is exploited to infer and to learn several parameters of the model for predicting the attributes of the test target samples. The experimental results from the subset of H3D dataset and VOC2011 dataset have shown that the proposed method is feasible and effective even if the training sample set is small.

• 出版日期2013-9-20
• 单位长沙理工大学