テキストに基づく運動生成手法を用いた歩行姿勢推定のためのデータセット構築

抄録

In this study, we leveraged Text to Motion (T2M) technology for the facile construction of a dataset aimed at joint angle estimation and verified its validity. Our findings revealed that T2M data could be generated with relative ease, devoid of environmental constraints such as force plates, thereby allowing for the production of variably rich data. However, the current T2M dataset exhibits inferior estimation accuracy when compared to evaluations using AIST models on AIST data and T2M models on T2M data. Discrepancies in minor features between AIST and T2M data, particularly noticeable at endpoints, were identified as potential causes. Addressing these discrepancies by aligning the detailed characteristics of T2M (simulation data) and AIST (real data) closely emerges as a future challenge. We propose that combining datasets or altering learning methodologies, such as Fine-tuning, could be effective strategies for enhancing estimation accuracy by bridging the gap between these data groups.