High-precision 3D reconstruction of Small Celestial Bodies (SCBs) is crucial for deep space exploration missions. In this paper, an attention-based SCB reconstruction method, DS-MVSNet, is proposed to solve the problems of poor reconstruction accuracy and low reconstruction efficiency in some regions in the existing methods. Firstly, an attention mechanism model is used to improve the deep feature extraction ability of 2D convolution in feature sparse and light and dark weak regions. Secondly, depth estimation is achieved and optimized by constructing a cost volume. Finally, the 3D point cloud after depth fusion is generated. Meanwhile, this paper also establishes a 3D reconstruction dataset of small celestial bodies to provide data support for the autonomous intelligent algorithm. Compared with the traditional multi-view method, the method proposed in this study has higher accuracy and efficiency and can provide a benchmark for future probe landing navigation algorithms.