A method of human motion reconstruction with sparse joints based on attention mechanism

Motion capture has significant applications in fields such as rehabilitation, virtual reality, and beyond. The cost-effective utilization of sparse sensors is crucial. The central challenge lies in achieving full-body motion estimation with a reduced sensor count. This paper proposes a neural network model based on attention mechanism for motion reconstruction with sparse joints. Compared to traditional methods, our method places greater emphasis on the spatial characteristics of motion sequences by using spatial attention, thus having lower reconstruction error, while preserving high visual quality. Furthermore, our method employs two encoders, with each being responsible for motion feature extraction and motion reconstruction, respectively, thereby enhancing robustness when dealing with new data. Experiments also show that our method can handle missing markers problem with low error.