Deep Ordinal Regression Network for Monocular Depth Estimation (1806.02446v1)

Abstract: Monocular depth estimation, which plays a crucial role in understanding 3D scene geometry, is an ill-posed problem. Recent methods have gained significant improvement by exploring image-level information and hierarchical features from deep convolutional neural networks (DCNNs). These methods model depth estimation as a regression problem and train the regression networks by minimizing mean squared error, which suffers from slow convergence and unsatisfactory local solutions. Besides, existing depth estimation networks employ repeated spatial pooling operations, resulting in undesirable low-resolution feature maps. To obtain high-resolution depth maps, skip-connections or multi-layer deconvolution networks are required, which complicates network training and consumes much more computations. To eliminate or at least largely reduce these problems, we introduce a spacing-increasing discretization (SID) strategy to discretize depth and recast depth network learning as an ordinal regression problem. By training the network using an ordinary regression loss, our method achieves much higher accuracy and \dd{faster convergence in synch}. Furthermore, we adopt a multi-scale network structure which avoids unnecessary spatial pooling and captures multi-scale information in parallel. The method described in this paper achieves state-of-the-art results on four challenging benchmarks, i.e., KITTI [17], ScanNet [9], Make3D [50], and NYU Depth v2 [42], and win the 1st prize in Robust Vision Challenge 2018. Code has been made available at: https://github.com/hufu6371/DORN.

• The paper presents a novel ordinal regression framework that reformulates depth estimation into a discrete task using SID, enhancing training convergence.
• It employs a multi-scale network architecture with atrous spatial pyramid pooling to efficiently capture rich spatial context without losing resolution.
• Evaluations on benchmarks like KITTI and ScanNet demonstrate state-of-the-art accuracy and robust performance across diverse depth ranges.

Deep Ordinal Regression Network for Monocular Depth Estimation

The paper "Deep Ordinal Regression Network for Monocular Depth Estimation" introduces an innovative method for monocular depth estimation (MDE) from a single image. This ill-posed problem is crucial for applications involving 3D scene understanding such as object recognition, segmentation, and detection. The solution proposed aims to mitigate issues present in previous methods that modeled MDE as a regression problem and employed deep convolutional neural networks (DCNNs).

Methodology

Key Strategies and Design

1. Discretization with SID: The paper employs a spacing-increasing discretization (SID) strategy to transform continuous depth values into discrete intervals. This method leverages the insight that uncertainty in depth estimation grows with increasing depth values. Discretizing depth helps in simplifying the training process and enhancing convergence speed.
2. Ordinal Regression Framework: Depth estimation is transformed from a standard regression task into an ordinal regression problem. This involves training the network with a loss function that respects the ordinal nature of depth values, yielding much better convergence and final accuracy.
3. Multi-Scale Network Architecture: The architecture avoids traditional spatial pooling, which typically leads to low-resolution feature maps. Instead, it uses a multi-scale network to capture spatial information efficiently. This is achieved through atrous spatial pyramid pooling (ASPP) and incorporation of dilated convolutions, ensuring large receptive fields without degrading resolution.
4. Full-Image Encoder: The full-image encoder captures global contextual information crucial for accurate depth prediction. This encoder uses fewer parameters compared to traditional fully-connected layers, reducing computational cost and memory usage while maintaining high performance.

Performance and Evaluation

The proposed model, which is benchmarked against several challenging datasets (KITTI, Make3D, NYU Depth v2, and ScanNet), has led to state-of-the-art results. Some notable performance metrics include:

• KITTI Benchmark:

The model achieves high accuracy, outperforming several previous methods even on challenging depth ranges (e.g., 0-80m).

• Make3D and NYU Depth v2:

Significant improvements are reported in usual evaluation metrics, supporting the robustness of the method across varied dataset conditions.

• ScanNet:

On the ScanNet dataset, the method ranks first in the Robust Vision Challenge, further underscoring its effectiveness.

Implications and Future Work

The innovative approach to MDE using SID and ordinal regression frameworks addresses significant challenges in the field, such as slow convergence and the complexity of network architectures. By discretizing depth into intervals and treating it as an ordinal regression task, the proposed method not only achieves higher accuracy but also ensures faster convergence.

The practical implications are vast, with potential applications spanning autonomous driving, augmented reality, and robotics. Theoretically, this approach prompts further exploration into discretization strategies and their role in various depth estimation tasks.

Future work could extend to other dense prediction problems, investigating new approximations and extending the efficient learning framework introduced here. Additionally, there is room for improvement in reducing computational requirements further while maintaining or enhancing performance metrics.

Conclusion

The Deep Ordinal Regression Network for Monocular Depth Estimation presents a solid advancement in the domain of depth estimation. By addressing inefficiencies in existing methods and introducing a novel architecture and training strategy, the research sets a new benchmark for accuracy and efficiency in monocular depth estimation tasks.