Despite the broad applications of Human Mesh Recovery (HMR), existing methods face critical limitations: an inability to predict metric-scale SMPL coordinates, poor robustness under occlusion, and a lack of effective integration with advancements in single-view HMR. Therefore, we introduce FMV-HMR (Flexible Multi-View Human Mesh), which predicts the metric depth—and hence the metric coordinates—of the SMPL model via triangulation. Experiments demonstrate that our method surpasses state-of-the-art absolute depth prediction approaches in metric depth estimation. By further incorporating occlusion weights and spatial weights, it boosts the accuracy of the fused SMPL landmarks. Experimental results demonstrate that our model surpasses state-of-the-art methods in estimating SMPL on both the Human3.6M and MPI-INF-3DHP datasets. Moreover, experiments conducted on datasets with added occlusion confirm the model’s effectiveness in mitigating the impact of occlusion. Moreover, the proposed framework can plug in virtually any off-the-shelf single-view model as its backbone, build an end-to-end pipeline, and be fine-tuned to attain state-of-the-art performance.
Lin et al. (Sat,) studied this question.