Refining Pre-Existing Graph Convolutional Networks with Locked Transformers

In this paper, the authors explore the problem of predicting human motion using machine learning techniques. They highlight the challenges of addressing this issue, such as the complexity of human movements and the need for large amounts of training data. The authors then introduce their proposed method, which combines a graph convolutional network (GCN) with a multi-layer perceptron (MLP) to predict human motion. They demonstrate the effectiveness of their approach on two datasets: Human3.6M, a benchmark dataset for human motion prediction, and Use Case 2, a scenario where the goal is to refine preexisting GCNs for human motion prediction.
The authors provide a detailed explanation of their method, starting with the GCN component, which captures long-range dependencies in the graph structure of the human body. They then introduce the MLP component, which processes the node features and predicts the final motion. The authors also discuss the importance of the "back to MLP" step, which helps to improve the performance of their method by refining the GCN predictions with a simple MLP.
The authors evaluate their method on both datasets and show that it outperforms existing approaches in terms of accuracy and efficiency. They also provide a thorough analysis of their method’s strengths and limitations, highlighting areas for future improvement.
Overall, this paper makes an important contribution to the field of human motion prediction by proposing a novel approach that combines the strengths of GCNs and MLPs. The authors provide a clear and concise explanation of their method, making it accessible to readers with varying levels of expertise in machine learning and computer vision.