EVENT DOSSIER
The research team proposed the SoundMHPE method for the first time, aiming to estimate multi-person 3D poses using only acoustic signals. The framework consists of two core components: the Acoustic Multi-scale Encoder, which separates subtle features from complex overlapping acoustic signals; and the Temporal Pose Decoder, which decouples cross-frame information through an attention mechanism to reconstruct individual poses. To verify the effectiveness of this method, the researchers constructed an AMP dataset containing 432,000 frames of synchronized data. Experimental results show that SoundMHPE outperforms existing baseline models in performance.
本文提出 SoundMHPE，一种仅利用声学信号估计多人 3D 姿态的首次尝试。该框架包含两个核心组件：Acoustic Multi-scale Encoder 用于从复杂重叠信号中分离细微声学特征，Temporal Pose Decoder 则通过注意力机制解耦跨帧的多人体信息以重建个体姿态。为验证该方法，研究者构建了包含 43.2 万帧同步数据的 AMP 数据集，实验表明 SoundMHPE 优于基线模型。