Robust 3D Human Pose Estimation with Unsynchronized Cross-View Fusion

Multi-View 3D human pose estimation tasks typically require fully synchronized input images. However, this strict synchronization requirement brings two significant challenges: (i) increased complexity and cost of data acquisition and (ii) reliance on pre-synchronized inputs during inference, limiting practical applicability. To address these issues, this paper introduces a novel approach for 3D human pose estimation based on unsynchronized input. This algorithmic advancement significantly reduces the complexity and costs of data acquisition while improving the robustness of the model during inference. In this new setup, the model must overcome errors caused by unsynchronized inputs and achieve effective convergence. We propose a Cross-View Fusion (CVF) module for cross-view correction to address the mistakes introduced by unsynchronized views. Additionally, we design a Slow Convergence Loss (SCL) to smooth model convergence during training. Extensive experiments on the Human3.6M dataset demonstrate the superior performance of our method.