A 3D reconstruction model based on attention and intermediate fusion representation is proposed, aiming to reconstruct a 3D model with refined structure. This method used the axial spatial attention mechanism to learn information in different directions, and embeded it in the encoder to capture local structural features. Then, based on the two-stream network, the depth map and 3D average shape could be inferred to design an intermediate fusion representation module which could effectively fuse visible surface details to better describe the 3D spatial structure of the objects. The experimental results show that the axial spatial attention mechanism and the intermediate fusion representation module proposed in this paper enhance the ability of feature extraction. The IoU and F-score are 1.3% and 0.4% higher than PixVox++, respectively, proving that the 3D reconstruction effect is better.