基于图卷积网络的三维手部姿态估计OA

In the task of 3D hand pose estimation from a single image in color,challenges such as occlusion and high self-similarity of hand parts might lead to large prediction errors and unnatural hand structures.To address these issues,a graph convolution-based 3D hand pose estimation method was firstly proposed.Visual features and 2D keypoint positions were extracted from the input image using Keypoint R-CNN.These features were then fed into an improved adaptive kernel graph convolution module(AK_GraFormer).Subsequently,a residual-connected AKGNN graph kernel was introduced to adaptively process graph-structured data,thereby enhancing the model's feature learning and representation.Finally,a dynamic training strategy was employed,which was monitored by a proposed evaluation metric,to optimize estimation performance.Experimental results on the HO-3D_v3 and Frei-Hand datasets demonstrated that the proposed method outperformsed existing approaches in monocular 3D hand pose estimation.Specifically,the procrustes-aligned mean per joint position error(PA-MPJPE)was reduced by up to 12.50 percentage points,and the area under the curve(AUC)of the percentage of correct keypoints metric was improved by up to 3.44 percentage points compared to state-of-the-art methods.