Scholarlink

<p>Unlike Transformers, multi-layer perceptrons with gating employ a simple gating mechanism that efficiently captures dependencies without incurring the quadratic complexity associated with self-attention. This paper introduces SimGate, a simple gated network for 3D human motion prediction. The proposed model employs the discrete cosine transform to encode motion information in the frequency domain, spatial and frequency projections to capture spatial relationships within individual frames and temporal dependencies across frames of the input motion, a gating mechanism that selectively controls the flow of information between these projections, and a prediction layer to generate the predicted 3D human poses for future time steps. Experimental evaluations conducted on three benchmark datasets demonstrate that SimGate achieves competitive performance relative to strong baselines in both quantitative and qualitative comparisons, surpassing the best-performing baseline with a relative average error reduction of 0.4%. Moreover, the results highlight SimGate’s ability to generate realistic motion sequences that closely align with ground truth, even in challenging scenarios involving occlusions. In addition, SimGate maintains a compact model size, making it lightweight and computationally efficient.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SimGate: a simple gated network for 3D human motion prediction

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Ryan Amstutz,
  • A. Ben Hamza

摘要

Unlike Transformers, multi-layer perceptrons with gating employ a simple gating mechanism that efficiently captures dependencies without incurring the quadratic complexity associated with self-attention. This paper introduces SimGate, a simple gated network for 3D human motion prediction. The proposed model employs the discrete cosine transform to encode motion information in the frequency domain, spatial and frequency projections to capture spatial relationships within individual frames and temporal dependencies across frames of the input motion, a gating mechanism that selectively controls the flow of information between these projections, and a prediction layer to generate the predicted 3D human poses for future time steps. Experimental evaluations conducted on three benchmark datasets demonstrate that SimGate achieves competitive performance relative to strong baselines in both quantitative and qualitative comparisons, surpassing the best-performing baseline with a relative average error reduction of 0.4%. Moreover, the results highlight SimGate’s ability to generate realistic motion sequences that closely align with ground truth, even in challenging scenarios involving occlusions. In addition, SimGate maintains a compact model size, making it lightweight and computationally efficient.