Scholarlink

Remote photoplethysmography (rPPG) has shown promise for non-contact heart rate estimation and face anti-spoofing. However, existing methods often suffer from a performance trade-off between motion robustness and spoof discriminability, and neglect the guiding role of facial structure. This paper presents TrackPhys, a novel unified network that leverages identity-guided, cross-task representation learning to simultaneously address both challenges. TrackPhys innovatively integrates facial contour information to guide physiological signal extraction, enhancing robustness against complex motions. Furthermore, the rPPG feature maps derived from our model are leveraged for highly discriminative spoof detection. Extensive experiments demonstrate state-of-the-art performance: TrackPhys reduces the mean absolute error (MAE) of heart rate estimation by 23% under extreme motion on cross-dataset tests. When integrated into a spoof detection framework, it achieves an exceptional EER of 0.62% on the 3DMAD dataset for long-term evaluation and 2.39% for challenging 1-second short-term prediction, significantly outperforming all previous methods.

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

TrackPhys: Learning Transferable Physiological Representations for Motion-Robust Heart Rate and 3D Mask Attack Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wenli Zhang,
  • Dali Zhu,
  • Hualin Zeng,
  • Long Yang

摘要

Remote photoplethysmography (rPPG) has shown promise for non-contact heart rate estimation and face anti-spoofing. However, existing methods often suffer from a performance trade-off between motion robustness and spoof discriminability, and neglect the guiding role of facial structure. This paper presents TrackPhys, a novel unified network that leverages identity-guided, cross-task representation learning to simultaneously address both challenges. TrackPhys innovatively integrates facial contour information to guide physiological signal extraction, enhancing robustness against complex motions. Furthermore, the rPPG feature maps derived from our model are leveraged for highly discriminative spoof detection. Extensive experiments demonstrate state-of-the-art performance: TrackPhys reduces the mean absolute error (MAE) of heart rate estimation by 23% under extreme motion on cross-dataset tests. When integrated into a spoof detection framework, it achieves an exceptional EER of 0.62% on the 3DMAD dataset for long-term evaluation and 2.39% for challenging 1-second short-term prediction, significantly outperforming all previous methods.