A Joint Network for Pose Classification and Evaluation Based on Attention Mechanism

In recent years, many studies have focused on pose recognition, with few studies on pose evaluation and even fewer on pose classification. Thus, studies combining pose evaluation and pose classification are almost non-existent. In this paper, a channel attention-based Siamese network was proposed, and two functions of pose classification and pose evaluation were integrated into the same network, which means this proposed model simultaneously implements classifying input poses into the appropriate categories and evaluating the similarity of input poses. Briefly, the proposed model was structured as a basic Siamese network with three convolutional blocks, one of which contains a convolution layer, a pooling layer, a batch normalization layer, and an activation function. After three convolutional blocks, one of the branches of the Siamese network diverged into two fully connected layers to generate two feature vectors, one for classification and the other for evaluation. The other branch connects a fully connected layer to generate a feature vector for pose evaluation. Furthermore, the attention module has been shown to be effective according to our experiments on the Baduanjin dataset. Experimental results also indicate that the performance of both the pose classification task and the pose evaluation task is relatively great in our proposed model.