Graph Embedded Representation Learning in Skeleton-based Action Classification

The wide usage of GCN and GNN on the tasks of action classification tasks has made great improvement since the first proposed ST-GCN model. Plenty of works proposed methodologies based on the classification task requirements. As a result of them, we manipulated similar skeleton structures which are extracted from images and videos by analyzing intra and inter-class to represent the behaviour isomerism graphs. Our methodology manipulated the unsupervised graph embedding methodology to solve the classification downstream tasks based on the collected large-scale 2-dimensional datasets. We apply our proposed methodology on top of 4 pose estimation datasets to verify the effectiveness of the results. To solve the unsuper-vised classification problem, we are focusing on the property of skeleton data which is view-invariant through manipulating the attention-based and encoder-decoder structure to generate the corresponding embeddings and compare them through the contrastive learning methodology.