Deep Ordinal Regression Network for Monocular Depth Estimation

By Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, Dacheng Tao. 2018

2,009 citations recorded by the source corpus.

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 multilayer 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 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 proposed deep ordinal regression network (DORN) achieves state-of-the-art results on three challenging benchmarks, i.e., KITTI [16], Make3D [49], and NYU Depth v2 [41], and outperforms existing methods by a large margin.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer (2019, methodology): “For DORN [19] we followed their tiling protocol, resizing the images to the dimensions stated for their NYU and KITTI evaluation, respectively.”
Vision Transformers for Dense Prediction (2021, neutral): “25 3 AbsRel RMSE log10 DORN [13] Zero-shot cross-dataset transfer.”
DIODE: A Dense Indoor and Outdoor DEpth Dataset (2019, background): “Make3D [27] was an early approach that leveraged supervised learning for monocular depth estimation, and more recent work has applied deep neural networks to the task [8, 18, 26, 22, 11, 11].”
From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation (2019, methodology): “[13] achieves state-of-the-art result by also taking advantages from changing the regression problem to quantized ordinal regression.”
UniDepth: Universal Monocular Metric Depth Estimation (2024, methodology): “While existing MMDE methods [2, 13, 15, 33, 34, 36, 50] have demonstrated remarkable accuracy across different benchmarks, they require training and testing on datasets with similar camera intrinsics and scene scales.”
Digging Into Self-Supervised Monocular Depth Estimation (2018, background): “Various approaches, such as combining local predictions [19, 55], non-parametric scene sampling [24], through to end-to-end supervised learning [9, 31, 10] have been explored.”
D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry (2020, background): “Fu et al. [24] recast the monocular depth estimation network as an ordinal regression problem and achieve superior performance.”
Pattern-Affinitive Propagation Across Depth, Surface Normal and Semantic Segmentation (2019, methodology): “Compared with the state-of-theart methods, though slightly weaker than DORN [15], our method obtains superior performances than all other published or unpublished approaches.”
Dual Attention Guided Gaze Target Detection in the Wild (2021, background): “Recently, CNN based approaches [11, 25, 15, 26] show significant improvements in this field.”
Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation (2022, methodology): “With the annotated depth one can train a regression CNN to predict the depth value of each pixel on a single image [10, 11, 22].”
Monocular depth estimation based on deep learning: An overview (2020, methodology): “[60] regard the monocular depth estimation as an ordinal”
Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving (2019, background): “According to (Wang et al., 2019a), depth estimation from stereo pairs of images (Mayer et al., 2016; Yamaguchi et al., 2014; Chang & Chen, 2018) are more accurate than that from monocular (i.e., single) images (Fu et al., 2018; Godard et al., 2017) for 3D object detection.”
Self-Supervised Monocular Depth Hints (2019, background): “The clear winner here is DORN [6], who avoid selfsupervision entirely, training directly from LiDAR data.”
Atlas: End-to-End 3D Scene Reconstruction from Posed Images (2020, methodology): “Another approach to 3D reconstruction is to use monocular [18,31,32], binocular [3,5] or multivew [23,27,28,51] stereo methods which take RGB images (one, ar X iv :2 00 3.”
FastDepth: Fast Monocular Depth Estimation on Embedded Systems (2019, background): “Since then, depth estimation has evolved from using simple hand-crafted feature representations [6] to modern deep learning based approaches [11–14].”
Enforcing Geometric Constraints of Virtual Normal for Depth Prediction (2019, methodology): “In addition to the quantitative comparison, we demonstrate some visual results between our method and the stateof-the-art DORN in Fig.”
SpatialVLM: Endowing Vision-Language Models with Spatial Reasoning Capabilities (2024, methodology): “Spatial distance estimation has been traditionally addressed as a part of broader tasks, such as SLAM [8, 21] or depth estimation [23].”