Abstract: Aiming at the problem of spatial and feature structure loss in the feature upsampling of decoder for deep learning model on point cloud, an attention-based decoder for feature interpolation upsampling was proposed. The decoder mainly consists of spatial attention and feature attention feature interpolation upsampling. First, the spatial interpolated features of the upsampled points were calculated by spatial attention interpolation based on the attention relationship between the spatial location features of the points, then the feature interpolation was calculated from the feature attention relationship between the spatial interpolated features and the subsampling points, finally the interpolated features were skip connected with the encoder features and the decoder features were output after feature mix. In the experiments, the proposed decoder was embedded into various deep learning models on point clouds, and the effectiveness of the decoder was verified on both ShapeNet and ModelNet datasets. The experimental results show that the proposed method improved the average overlap by 0.5% in the part segmentation task and reduced the average cosine distance error by 13% in the normal vector estimation task compared with the traditional three-point linear interpolation model, which fully verified the effectiveness of the decoder.