基于隐式表示的全局-局部关系注意力三维重建方法研究OA

The implicit representation 3D reconstruction method is an innovative technique oriented to efficiently reconstruct complex 3D models in the field of computer vision,which has important practical application value.The existing 3D reconstruction techniques face problems such as geometric distortion and poor detail recovery,especially in the reconstruction of sparse point clouds and complex geo-metric structures.For this reason,a global-local relational attention 3D reconstruction method based on implicit representation is pro-posed,which firstly extracts point and mesh features by using the point-mesh Transformer module,and then comprehensively captures the complex relationship between point features and mesh features through the global-local relational attention module,which not only considers the local geometric information when updating the features,but also incorporates the global structural information.Meanwhile,a multi-head codebook strategy is introduced to establish a priori knowledge across categories,which improves the expressiveness of the model in reconstructing objects of different categories.Finally,an edge-based marching cube algorithm is used for reconstruction,which solves the boundary blurring problem of existing methods in sharp edges and complex structures.It is experimentally demonstrated that the proposed method achieves significant improvement in the evaluation metrics IoU,CD and F-Score on the ShapeNet dataset.