Efficient Material Editing of Single Image Based on Inverse Rendering

Fig. 6. Visual comparisons between the synthetic images (GT) after the combination of specular layer and diffuse layer with different material shininess parameters and the output results of the proposed method

Download full size

Fig. 7. Material editing of real pictures

Download full size

Table 1. Hyperparameter settings in network training
View table
Table 1. Hyperparameter settings in network training
Network Batch size Optimizer Initial learning rate Number of iterations
Specular-Net 64 Adam 1×10^-4 50×10³
IID-Net 64 Adam 1×10^-4 28×10³
Shininess-Net 64 Adam 1×10^-4 35×10³

Table 2. Datasets used by proposed networks
View table
Table 2. Datasets used by proposed networks
Network Dataset Source Size
Specular-Net LIME Meka et al.^［4］ 8.5×10⁴
IID-Net ShapeNet Intrinsics Shi et al.^［5］ 20×10⁴
Shininess-Net Hierarchical Shininess Ours 140×10⁴

Table 3. Quantitative comparisons on the SHIQ dataset, ShapeNet Intrinsics dataset, and some real pictures

View table

Table 3. Quantitative comparisons on the SHIQ dataset, ShapeNet Intrinsics dataset, and some real pictures

Method	SHIQ dataset			ShapeNet Intrinsics dataset			Real pictures
Method	SMSE	LMSE	DSSIM	SMSE	LMSE	DSSIM	SMSE	LMSE	DSSIM
Shen13^［28］	0.0587	0.333	0.2002	0.1286	0.2433	0.207	0.0224	0.167	0.1745
Shi17^［5］	0.0305	0.1789	0.2199	0.0133	0.1056	0.1576	0.0167	0.2497	0.2087
Souza18^［9］	0.0607	0.349	0.2062	0.1258	0.2391	0.2103	0.009	0.1178	0.1639
Yamamoto19^［29］	0.0594	0.3364	0.2011	0.1286	0.2418	0.206	0.0225	0.1667	0.1743
Fu21^［10］	0.0003	0.0019	0.0088	0.0107	0.0332	0.1034	0.0056	0.1035	0.1414
Ours	0.0108	0.0485	0.1765	0.0078	0.0449	0.0927	0.0056	0.1448	0.1384

Table 4. Quantitative comparisons on the MIT intrinsics dataset

View table

Table 4. Quantitative comparisons on the MIT intrinsics dataset

Method	SMSE		LMSE		DSSIM
Method	albedo	shading	albedo	shading	albedo	shading
SIRFS15^［30］	0.0147	0.0083	0.0416	0.0168	0.1238	0.0985
DI15^［15］	0.0277	0.0154	0.0585	0.0295	0.1526	0.1328
Shi17^［5］	0.0278	0.0126	0.0503	0.0240	0.1465	0.1200
Yi20^［19］	0.0274	0.0145	0.0476	0.0284	‒	‒
SMCH21^［31］	0.0225	0.0146	0.0484	0.0278	0.1499	0.1912
Ours	0.0273	0.0129	0.0386	0.0335	0.1484	0.1420

Tools

Get Citation

Copy Citation Text

Kunliang Xie, Renjiao Yi, Haifang Zhou, Chenyang Zhu, Yuwan Liu, Kai Xu. Efficient Material Editing of Single Image Based on Inverse Rendering[J]. Laser & Optoelectronics Progress, 2022, 59(14): 1415014

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites