Constrained Multi-Stencil Least-Square Predictors for Triangle Mesh Compression

To improve the compression rate of triangle meshes,this paper proposes a data-driven multi-stencil least-square prediction method,making full use of the features extracted from local neighborhood information of the mesh.In the training phase,features from all the possible five-vertex stencils are used as the training set,and the predictor is constructed in the local coordinate system whose weights are obtained by the least square method.In the encoding phase,the prediction of the current vertex uses a constrained multi-stencil strategy,optimizing the linear combination of the available stencils,followed by entropy encoding on the residue.Unlike the mesh-independent prediction schemes(such as the parallelogram predictor),the method makes further use of the correlation between adjacent triangles on the mesh,effectively reducing the error of prediction,which improves the compression rate.With the same quantization error and traverse order,the proposed method can generally obtain higher and stable compression rate than existing prediction methods,especially on the smooth models.