Researching | Inverse Reflectance Model Based on Deep Learning

Acta Optica Sinica, Volume. 43, Issue 21, 2120001(2023)

Inverse Reflectance Model Based on Deep Learning

Xi Wang, Zhenxiong Jian, and Mingjun Ren^*

Author Affiliations

State Key Laboratory of Mechanical System and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (14)

References (38)

Paper Information

Figures & Tables(14)

Fig. 1. Inverse reflectance model based on deep learning

Download full size

Fig. 2. Represented images of synthetic testing dataset

Download full size

Fig. 3. Represented images of proposed supplementary training dataset

Download full size

Fig. 4. Qualitative results of synthetic experiment

Download full size

Fig. 5. Experiment results on sparse light scene

Download full size

Fig. 6. Qualitative experiment results of every stage subnetwork

Download full size

Fig. 7. Qualitative results of real experiments

Download full size

Table 1. Hyper-parameter settings of different training stages

Table 1. Hyper-parameter settings of different training stages

Hyper-parameter setting	Stage 1	Stage 2	Stage 3	Combine
Initial learning rate	0.0001	0.0001	0.001	$10^{- 6}$
Total epoch	40	40	40	5
Batch size	32	32	32	16
Learning rate decay epoch	［26，36］	［26，36］	［26，36］	-
Noise level	［-0.005，0.005］	［-0.005，0.005］	［-0.005，0.005］	［-0.01，0.01］

Table 2. Ablation experiment results
View table
Table 2. Ablation experiment results
Maxpooling Stage 1 MAVE
$Δ φ$ $n^{T} l$
× × 0.101 0.064
√ × 0.030 0.040
√ √ - 0.033

Table 3. MAE measured on dense light scene
View table
Table 3. MAE measured on dense light scene
Method MAE /（°） Average MAE /（°）
Armadillo Bunny Dragon Sphere
Ours 2.3 1.9 2.2 2.3 2.2
WJ20^［33］ 3.2 2.6 3.1 2.8 2.9
IK18^［20］ 5.1 4.1 5.6 3.6 4.6
CH20^［19］ 5.7 4.2 5.4 3.9 4.8

Table 4. MAE measured on SVBRDF scene
View table
Table 4. MAE measured on SVBRDF scene
Method MAE /（°） Average MAE /（°）
Armadillo Bunny Dragon Sphere
Ours 2.6 1.9 2.6 1.8 2.2
WJ20^［33］ 3.4 2.5 3.5 2.6 3.0
IK18^［20］ 5.1 3.8 5.8 3.6 4.6
CH20^［19］ 5.9 4.0 5.6 3.9 4.9

Table 5. Experiment results of every stage subnetwork

Table 5. Experiment results of every stage subnetwork

Object	Ball	Bear	Budd.	Cat	Cow	Gobl.	Harv.	Pot1	Pot2	Read.	Aver.
MAVE in stage 1	0.011	0.020	0.023	0.020	0.025	0.022	0.046	0.020	0.022	0.030	0.024
MAE in stage 2 /（°）	1.7	4.2	6.7	4.1	5.5	7.2	12.8	5.2	6.0	10.3	6.37
MAE in stage 3 /（°）	2.0	4.0	6.3	4.0	5.0	6.6	11.7	4.9	5.8	9.9	6.02
MAE of combine /（°）	2.0	3.6	6.3	3.9	5.1	6.5	11.6	5.0	5.2	9.8	5.90

Table 6. Comparison results of normal vector on DiLiGenT dataset using all the images

Table 6. Comparison results of normal vector on DiLiGenT dataset using all the images

Method	Normal vector estimation error /（°）										Average error /（°）
Method	Ball	Bear	Budd.	Cat	Cow	Gobl.	Harv.	Pot1	Pot2	Read.	Average error /（°）
Ours	2.0	3.6	6.3	3.9	5.1	6.5	11.6	5.0	5.2	9.8	5.90
FI21^［28］	2.0	3.5	7.6	4.3	4.7	6.7	13.3	4.9	5.0	9.8	6.17
WJ20^［33］	1.6	4.6	6.9	4.7	5.2	7.8	13.0	5.6	6.6	10.2	6.62
IK18^［20］	2.2	4.1	7.9	4.6	8.0	7.3	14.0	5.4	6.0	12.6	7.21
CH20^［40］	2.4	4.6	7.5	4.8	8.0	8.9	13.9	5.7	6.7	10.5	7.30

Table 7. Comparison results of normal vector on DiLiGenT using 10 images

Table 7. Comparison results of normal vector on DiLiGenT using 10 images

Method	Normal vector estimation error /（°）										Average error /（°）
Method	Ball	Bear	Budd.	Cat	Cow	Gobl.	Harv.	Pot1	Pot2	Read.	Average error /（°）
Ours	2.1	4.3	7.2	5.2	6.2	8.3	13.8	6.1	6.7	10.7	7.06
WJ20^［33］	2.3	5.9	7.9	5.9	7.2	8.9	15.5	6.7	8.2	11.3	7.98
CH20^［19］	2.9	5.9	8.7	5.4	9.1	10.6	15.6	6.5	8.3	10.6	8.36
FI21^［28］	2.5	4.9	9.4	6.3	7.2	9.7	16.1	7.0	7.7	13.1	8.37

Tools

Get Citation

Copy Citation Text

Xi Wang, Zhenxiong Jian, Mingjun Ren. Inverse Reflectance Model Based on Deep Learning[J]. Acta Optica Sinica, 2023, 43(21): 2120001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Optics in Computing

Received: Mar. 2, 2023

Accepted: Jun. 13, 2023

Published Online: Nov. 16, 2023

The Author Email: Ren Mingjun (renmj@sjtu.edu.cn)

DOI:10.3788/AOS230615

Topics

laser devices and laser physics

Lasers and Laser Optics

laser manufacturing

Instrumentation, Measurement and Metrology