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.

Hierarchical Quadruplet Net for Deep Metric Learning and Network Regularization

Learning Local Feature Descriptors with Quadruplet Ranking Loss

Discriminatively Learning for Representing Local Image Features with Quadruplet Model

Deep Quadruplet Network For Hyperspectral Image Classification With A Small Number Of Samples

SSML-QNet: Scale-Separative Metric Learning Quadruplet Network for Multi-Modal Image Patch Matching

A Deep Quadruplet Network for Local Descriptor Learning

A Hierarchical Sampling Based Triplet Network for Fine-Grained Image Classification

HCFNN: High-order Coverage Function Neural Network for Image Classification

Superpixel-Based Long-Range Dependent Network for High-Resolution Remote-Sensing Image Classification

Deep Metric Learning with Hierarchical Triplet Loss.

Deep Ordinal Regression Network for Monocular Depth Estimation

Self-Distilled Hierarchical Network for Unsupervised Deformable Image Registration

Deep Image Similarity Measurement Based on the Improved Triplet Network with Spatial Pyramid Pooling.

Deep quantum networks for classification

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

Deep Multi-task Learning for Hierarchical Classification

Deep Quantization Network for Efficient Image Retrieval

HDMNet: A Hierarchical Matching Network with Double Attention for Large-scale Outdoor LiDAR Point Cloud Registration

Deep Cube-Pair Network for Hyperspectral Imagery Classification

LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks