Abstract: To reduce the dependency of dense depth data, we propose a partial-order-relationship based depth estimation method, which learns monocular depth maps from the sparse 3 D point cloud. Firstly, the RGB image is divided into superpixels, and the depth values of superpixels are generated according to the projection of the sparse point cloud on the superpixel image. Then a set of depth partial order relationship is generated at the center of superpixels. Subsequently, the depth partial order relationship of the RGB image and the actual depth value of the sparse point cloud are combined to train a convolutional neural network to estimate the depth map of the scene. The experimental results on the sparse point clouds sampled from NYU Depth v2 show that our method achieves an average relative error of 0.262 with sparse point cloud, which is better than the conventional methods, and even better than some supervised methods using dense depth data.