This paper proposes a deep ordinal regression network (DORN) for monocular depth estimation. Monocular depth estimation is an ill-posed problem, as a single 2D image can correspond to an infinite number of 3D scenes. Existing methods often use regression to estimate depth, but this approach suffers from slow convergence and unsatisfactory local solutions. Additionally, depth estimation networks often use repeated spatial pooling, leading to low-resolution feature maps and requiring complex training procedures. To address these issues, the authors introduce a spacing-increasing discretization (SID) strategy to discretize depth values, transforming the depth estimation problem into an ordinal regression problem. This approach allows for more accurate depth prediction by considering the ordering of discrete depth values. The network architecture avoids unnecessary spatial pooling and captures multi-scale information in a simpler way, using a multi-scale feature learner and a full-image encoder. The proposed method achieves state-of-the-art results on four challenging benchmarks: KITTI, ScanNet, Make3D, and NYU Depth v2. It outperforms recent algorithms by a significant margin. The method is trained in an end-to-end manner without stage-wise training or iterative refinement. The network uses an ordinal regression loss to learn the parameters, which takes into account the ordering of discrete depth values. The method also uses a full-image encoder that captures global contextual information efficiently, reducing the number of parameters compared to traditional full-image encoders. The results show that the proposed method achieves higher accuracy and faster convergence compared to existing methods. The method is robust to different depth ranges and performs well on both indoor and outdoor data. The authors conclude that the proposed method is effective for monocular depth estimation and can be extended to other dense prediction problems.This paper proposes a deep ordinal regression network (DORN) for monocular depth estimation. Monocular depth estimation is an ill-posed problem, as a single 2D image can correspond to an infinite number of 3D scenes. Existing methods often use regression to estimate depth, but this approach suffers from slow convergence and unsatisfactory local solutions. Additionally, depth estimation networks often use repeated spatial pooling, leading to low-resolution feature maps and requiring complex training procedures. To address these issues, the authors introduce a spacing-increasing discretization (SID) strategy to discretize depth values, transforming the depth estimation problem into an ordinal regression problem. This approach allows for more accurate depth prediction by considering the ordering of discrete depth values. The network architecture avoids unnecessary spatial pooling and captures multi-scale information in a simpler way, using a multi-scale feature learner and a full-image encoder. The proposed method achieves state-of-the-art results on four challenging benchmarks: KITTI, ScanNet, Make3D, and NYU Depth v2. It outperforms recent algorithms by a significant margin. The method is trained in an end-to-end manner without stage-wise training or iterative refinement. The network uses an ordinal regression loss to learn the parameters, which takes into account the ordering of discrete depth values. The method also uses a full-image encoder that captures global contextual information efficiently, reducing the number of parameters compared to traditional full-image encoders. The results show that the proposed method achieves higher accuracy and faster convergence compared to existing methods. The method is robust to different depth ranges and performs well on both indoor and outdoor data. The authors conclude that the proposed method is effective for monocular depth estimation and can be extended to other dense prediction problems.