Abstract: Depth images completion aims to recover dense depth images from sparse depth images, however, the depth images restored by many current depth completion algorithms often suffer from problems such as missing detail structures, depth discontinuities and blurred boundaries. To address these problems, this paper proposes a sparse depth completion method based on boundary prediction assistance, in which depth image completion is the main task and boundary prediction is the auxiliary task. A cross-guidance module is proposed to realize information interaction between the main task and the auxiliary task and provide effective boundary constraint for the main completion task. Moreover, an intermediate feature extraction module is used to extract multiple perceptual field features for scene context learning. In this paper, a series of experiments are conducted on the indoor dataset NYUv2 and the outdoor dataset KITTI, the experimental results prove the effectiveness of the proposed algorithms and modules, and it is superior to some mainstream depth completion methods in qualitative and quantitative comparison.