Robust 3D Human Mesh Recovery with Silhouette-Assisted Joint Attention Network

3D human mesh recovery algorithm based on regressing statistic model parameters has the advantages of simple parameter form and high accuracy, making it the mainstream algorithm at present. However, it still has the following problems. First, in complex scenarios, the estimation results of the contact surfaces between the limbs and the external environment are particularly inaccurate, which can adversely affect downstream tasks such as action analysis and human-robot interaction. Second, when predicting limb pose, insufficient utilization of local features associated with limbs increases the difficulty of network learning. To overcome the above problems, we propose a robust 3D human mesh recovery with Silhouette-Assisted Joint Attention Network(SAJAN). First, the background-invariant silhouette feature auxiliary module is proposed, which utilizes precise and detailed human silhouette features to provide valuable information for estimating contact surfaces while effectively avoiding background interference. Second, we propose a joint attention guidance-based regression network for 3D human mesh recovery to address the issue of insufficient utilization of local features associated with each limb in predicting limb pose, enabling the network to fully utilize local features of the human body in the image. Our method is evaluated on benchmark datasets and exhibits competitive results.