Field Converter：结合几何初值与时序残差的足球转播全局三维人体姿态估计

Translation Process for Computer Vision Paper Summary

Okay, here's how I approached this. First, I carefully analyzed the request: a translation of a computer vision paper abstract, from English to Simplified Chinese. The key was to maintain a professional tone, preserve paragraph breaks, keep the original URLs (though there weren't any in this text), and focus solely on the translation. My thinking proceeded sentence by sentence. I broke down each sentence, identifying key technical terms like "3D human pose," "monocular sports broadcasts," "shared metric world coordinate system," and so on. I had to choose the most accurate and commonly used Chinese equivalents in the field, making sure to consider the nuances in terminology. For example, "recovering" versus "estimating" 3D human pose (两者在 CV 领域常用 "估计" 或 "重建") needed careful consideration. I wanted the translation to sound natural and precise. I decided to keep the name "Field Converter" as a proper noun. I crafted initial drafts for each sentence, and then refined them, paying close attention to phrasing and clarity. I also made sure to use consistent terminology throughout. Lastly, I reviewed the entire translated paragraph for fluency, consistency, and a professional tone, ensuring it read well in Chinese without any unnecessary commentary or extraneous information.

当需要将球员定位在统一的度量世界坐标系中，而不仅是相对于自身躯干进行重建时，从单目体育赛事转播中恢复三维人体姿态仍极具挑战。我们提出了 Field Converter，这是一种基于几何初始化的时序残差框架，用于从已标定的足球赛事转播中估计世界坐标系下的三维球员姿态。我们的方法首先利用相机和球场几何关系，通过射线与地面求交来初始化球员根节点，随后结合姿态、图像、相机以及几何线索预测时序残差修正量。在比赛互不重叠的评估序列上，残差细化将仅依靠几何方法时的 49 cm 根节点误差，在使用逐帧 MLP 时降至 14 cm，使用 TCN 时降至 10 cm，而 Transformer 也取得了相近的 11 cm。最终的世界空间 MPJPE 达到了 13.2 cm；消融实验表明，残差预测的效果显著优于直接回归全局根节点，且时序上下文比具体的时序主干网络更为关键。失败案例分析进一步表明，腾空运动是这种基于地面的几何初始化的主要局限所在。