Real-Time Indoor Layout Estimation Method Based on Multi-Task Supervised Learning

Rongze Huang; Qinghao Meng; Yinbo Liu

doi:10.3788/LOP202158.1410023

Laser & Optoelectronics Progress, Volume. 58, Issue 14, 1410023(2021)

Real-Time Indoor Layout Estimation Method Based on Multi-Task Supervised Learning

Rongze Huang, Qinghao Meng, and Yinbo Liu^*

School of Electrical and Information Engineering, Institute of Robotics and Autonomous Systems, Tianjin Key Laboratory of Process Detection and Control, Tianjin University, Tianjin 300072, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(8)

Fig. 1. General structure of multi-task supervised lightweight convolutional neural network

Download full size

Fig. 2. Structures of various convolution modules. (a) Non-bottleneck-1D; (b) LFBlock; (c) DSBlock; (d) USBlock

Download full size

Fig. 3. Examples of labels. (a) Original images; (b) edge annotation heat maps; (c) visualization result of semantic segmentation labels

Download full size

Fig. 4. Visualization results of the proposed network model. (a) Original images; (b) semantic segmentation ground truth maps; (c) semantic segmentation prediction maps of the proposed method; (d) comparison maps between the estimated layouts of the proposed method and the real layouts (green is the estimated layout, red is the real layout)

Download full size

Table 1. Parameters of the encoder

View table

Table 1. Parameters of the encoder

Layer ID	Block type	Dilation	Dropout	Output channels	Output resolution
1	DSBlock	--	--	16	128×128
2	DSBlock	--	--	64	64×64
3-7	LFBlock	1	0.3	64	64×64
8	DSBlock	--	--	128	32×32
9	LFBlock	2	0.3	128	32×32
10	LFBlock	4	0.3	128	32×32
11	LFBlock	8	0.3	128	32×32
12	LFBlock	16	0.3	128	32×32
13	LFBlock	2	0.3	128	32×32
14	LFBlock	4	0.3	128	32×32
15	LFBlock	8	0.3	128	32×32
16	LFBlock	16	0.3	128	32×32

Table 2. Parameters of the decoder
View table
Table 2. Parameters of the decoder
Layer ID Block type Dilation Dropout Output channels Output resolution
1 USBlock -- -- 64 64×64
2-3 LFBlock 1 0.3 64 64×64
4 USBlock -- -- 16 128×128
5-6 LFBlock 1 0.3 16 128×128
7 Deconvolution -- -- 1/15 256×256

Table 3. Model performance evaluation

View table

Table 3. Model performance evaluation

Addmulti-task supervised?	Use LFBLock?	MPA /%	MIOU /%	FWIOU /%	CE /%	PE /%	Size /MB	Time /ms
No	Yes	77.44	66.42	71.97	7.04	9.86	6.2	43.26
No	No	77.76	66.65	72.01	6.92	9.69	8.8	48.02
Yes	No	81.03	68.04	73.85	6.55	9.47	8.8	47.92
Yes	Yes	81.96	68.91	74.02	6.26	9.05	6.2	43.13

Table 4. Performance comparison of different methods on LSUN dataset
View table
Table 4. Performance comparison of different methods on LSUN dataset
Method CE /% PE/%
Ref. [4] 15.48 24.23
Ref. [6] 11.02 16.71
Ref. [22] 10.13 14.82
Ref. [23] 8.70 12.49
Ref. [10] 8.20 10.63
Ref. [24] 7.95 9.31
Ref. [9] 6.30 9.86
Proposed 6.26 9.05

Tools

Get Citation

Copy Citation Text

Rongze Huang, Qinghao Meng, Yinbo Liu. Real-Time Indoor Layout Estimation Method Based on Multi-Task Supervised Learning[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410023

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites