Monocular Depth Estimation Method Based on Plane Coefficient Representation with Adaptive Depth Distribution

Jiajun Wang; Yue Liu; Yuhui Wu; Hao Sha; Yongtian Wang

doi:10.3788/AOS230468

Acta Optica Sinica, Volume. 43, Issue 14, 1415001(2023)

Monocular Depth Estimation Method Based on Plane Coefficient Representation with Adaptive Depth Distribution

Jiajun Wang, Yue Liu^*, Yuhui Wu, Hao Sha, and Yongtian Wang

Beijing Engineering Research Center of Mixed Reality and Advanced Display, School of Optics and Photonics, Beijing Institute of Technology, Beijing 100081, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Schematic diagram of pixel depth distribution of different images in indoor scene. (a) Image with maximum depth unconstrained; (b) image with maximum depth constrained; (c) close-up image of furniture

Download full size

Fig. 2. Structure diagram of image depth distribution prediction module

Download full size

Fig. 3. Overall architecture of monocular depth estimation network. (a),(b) Subnetwork structure

Download full size

Fig. 4. Visual comparison of proposed method with other methods on NYU Depth-v2 dataset

Download full size

Fig. 5. Visual comparison of three-dimensional reconstruction of proposed method with other methods

Download full size

Fig. 6. Qualitative and quantitative comparison of depth distribution of proposed method with other methods

Download full size

Fig. 7. Failure predictions of proposed method on NYU Depth-v2 dataset

Download full size

Table 1. Quantitative comparison of proposed method with other methods on NYU Depth-v2 dataset

View table

Table 1. Quantitative comparison of proposed method with other methods on NYU Depth-v2 dataset

Method	RMSE	REL	lg	$δ < 1.25$	$δ < 1 . 25^{2}$	$δ < 1 . 25^{3}$	Params /M
Eigen et al.^［13］	0.641	0.158	—	0.769	0.950	0.988	141
Laina et al.^［42］	0.573	0.127	0.055	0.811	0.953	0.988	64
Hao et al.^［43］	0.555	0.127	0.053	0.841	0.966	0.991	60
Fu et al.^［44］	0.509	0.115	0.051	0.828	0.965	0.992	110
Hu et al.^［45］	0.530	0.115	0.050	0.866	0.975	0.993	157
Raman et al.^［46］	0.495	0.139	0.047	0.888	0.979	0.995	80.4
Lee et al.^［18］	0.419	0.119	0.051	0.865	0.975	0.993	49.5
Yin et al.^［28］	0.416	0.108	0.048	0.878	0.977	0.994	114.2
Proposed	0.416	0.121	0.050	0.864	0.974	0.995	46

Table 2. Inference speed comparison of proposed method with other methods

View table

Table 2. Inference speed comparison of proposed method with other methods

Method	Running time /ms	Frame rate /（frame· $s^{- 1}$ ）	RMSE	Params /M
Laina et al.^［42］ Fu et al.^［44］ Hu et al.^［45］	36 35 91	27 28 11	0.573 0.509 0.530	64 110 157
Raman et al.^［46］	21	41	0.495	80.4
Lee et al.^［18］ Yin et al.^［28］	60 36	16 27	0.419 0.416	49.5 114.2
Proposed	30	33	0.416	46

Table 3. Comparison of IoU results of proposed method with other methods on test set
View table
Table 3. Comparison of IoU results of proposed method with other methods on test set
Method D_IoU
Fu et al.^［44］ 0.863
Yin et al.^［28］ 0.805
Raman et al.^［46］ 0.883
Hu et al.^［45］ 0.887
Proposed 0.914

Table 4. Quantitative results of ablation experiments based on network architecture
View table
Table 4. Quantitative results of ablation experiments based on network architecture
Method RMSE REL lg $δ < 1.25$ $δ < 1 . 25^{2}$ $δ < 1 . 25^{3}$
Depth distribution predictor 0.430 0.127 0.052 0.853 0.971 0.994
Plane coefficient representations 0.420 0.123 0.051 0.862 0.975 0.995
Baseline 0.416 0.121 0.050 0.864 0.974 0.995

Table 5. Quantitative results of ablation experiments based on network loss function

View table

Table 5. Quantitative results of ablation experiments based on network loss function

Method	RMSE	REL	lg	$δ < 1.25$	$δ < 1 . 25^{2}$	$δ < 1 . 25^{3}$
Baseline	0.431	0.125	0.052	0.857	0.971	0.994
With $L_{e d g e}$	0.428	0.124	0.051	0.858	0.973	0.994
With $L_{e d g e}$ and $L_{v i r}$	0.426	0.123	0.051	0.859	0.973	0.994
With $L_{e d g e}$ and $L_{d i s}$	0.418	0.122	0.050	0.862	0.974	0.995
With $L_{e d g e}$ ， $L_{d i s}$ ，and $L_{v i r}$	0.416	0.121	0.050	0.864	0.974	0.995

Tools

Get Citation

Copy Citation Text

Jiajun Wang, Yue Liu, Yuhui Wu, Hao Sha, Yongtian Wang. Monocular Depth Estimation Method Based on Plane Coefficient Representation with Adaptive Depth Distribution[J]. Acta Optica Sinica, 2023, 43(14): 1415001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Jan. 12, 2023

Accepted: Mar. 20, 2023

Published Online: Jul. 13, 2023

The Author Email: Yue Liu (liuyue@bit.edu.cn)

DOI:10.3788/AOS230468

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Quantitative comparison of proposed method with other methods on NYU Depth-v2 dataset

Table 1. Quantitative comparison of proposed method with other methods on NYU Depth-v2 dataset

Table 2. Inference speed comparison of proposed method with other methods

Table 2. Inference speed comparison of proposed method with other methods

Table 3. Comparison of IoU results of proposed method with other methods on test set

Table 3. Comparison of IoU results of proposed method with other methods on test set

Table 4. Quantitative results of ablation experiments based on network architecture

Table 4. Quantitative results of ablation experiments based on network architecture

Table 5. Quantitative results of ablation experiments based on network loss function

Table 5. Quantitative results of ablation experiments based on network loss function