In this paper, we propose an intra prediction method using CNN and attention mechanism for coding high-resolution depth videos utilized in virtual reality. The proposed method enhances intra prediction performance for depth pictures by predicting spatial correlations between an input block and reference pixels which is adjacent to the block. The proposed network extracts spatial features through CNN layers and predicts the spatial correlations through attention mechanism. Spatial features in vertical and horizontal directions are extracted from top and left adjacent blocks, respectively, and merged to predict the spatial features of pixels in the input block. The attention layers predict correlations between the spatial features of the input block and the reference pixels. Finally, the pixel values are predicted through the predicted correlation. In the simulation results, the intra prediction accuracies are improved up to 3.37% compared with the intra modes of VVC.