Human pose evaluation based on full-domain convolution and LSTM

Abstract

Abstract In this paper, we first analyze full domain convolution and LSTM to evaluate human pose by convolutional neural network and LSTM network. Secondly, graph structure skeleton image and skeleton point image classifier based on CNN and LSTM is constructed. The two-dimensional pose assessment method and three-dimensional pose assessment method were used to empirically analyze the human pose assessment. The results show that the average accuracy mAP values of the traditional evaluation methods are 69.7, 72.3, 71.4, and 74.4, respectively, while the average accuracy mAP value of the method used for 2D pose evaluation is 74.6. Where the average error of LReLU is the smallest. This shows that full-domain convolution and LSTM can be effective for human pose evaluation.