With the rise of the application of point clouds,the research on the learning of dense correspondence be-tween 3D point clouds has received more attention.For the task of dense correspondence,the key is to improve the rep-resentation power and robustness of point-wise features of point clouds.However,the feature embedding module of the previous methods only aggregates features with local neighbors and directly concatenates each stage output as the final features.Therefore,a global feature fusion module is proposed to extract global context explicitly and integrate it into the local features.In addition,a multi-stage feature aggregation module is proposed to aggregate the features from different stages with attention mechanism.Extensive experiments on both human and animal datasets show that our method can make a performance boost in comparison to recent dense correspondence methods.
维普
万方数据
