三维人体姿态估计中均一化金字塔特征捕捉网络

Unified Pyramid Features Capture Network in 3D Human Pose Estimate

针对目前三维人体姿态估计较多采用的混合方法面对人体部件遗漏、肢体欠匹配等情况出现估计结果误差较大的问题,提出一种均一化金字塔特征捕捉网络。通过在支路中加入金字塔特征捕捉模块,采取多尺度特征提取方法,增强对语义、纹理和姿态细节特征的提取能力;将金字塔特征捕捉模块应用于该网络三条支路,组成均一化金字塔特征捕捉网络,使热力、2D和3D图像更加精准。在3DPW数据集上的实验结果表明,与三维人体姿态估计网络相比,人体姿态误差明显降低,MPJPE值下降1.9%,PA-MPJPE值下降3.1%。对VR、AR和动作捕捉等场景具有一定的应用价值。

Aiming at the hybrid methods mostly used in the current 3D human pose estimation,a unified pyramid feature capture network is proposed in the face of the problem that omission and low limb matching accuracy will cause the error of the estimation result to become larger.By adding a pyramid feature capture module to the branch,the network adopts a multi-level and multi-scale feature extraction method to enhance the ability to extract features of semantic,texture and pose details.Using the unifying processing method,the pyramid feature capture module is applied to three branches to form a unified pyramid feature capture network,which makes heat,2D and 3D images more accurate.The experimental results on the 3DPW dataset show that compared with the 3D human pose estimation network(ROMP),the human pose error is significantly reduced,the MPJPE value decreases by 1.9%,and the PA-MPJPE value decreases by 3.1%.It has certain application value for VR,AR,and other motion capture scenes.