Deep Graph Attention Convolution Network for Point Cloud Semantic Segmentation

Yujing Chai; Jie Ma; Hong Liu

doi:10.3788/LOP202158.1210016

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 12 >Page 1210016 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 12, 1210016 (2021)

Deep Graph Attention Convolution Network for Point Cloud Semantic Segmentation

Yujing Chai, Jie Ma^*, and Hong Liu

Author Affiliations

School of Electronic Information Engineering, Hebei University of Technology, Tianjin 300401, China

show less

DOI: 10.3788/LOP202158.1210016 Cite this Article Set citation alerts

Yujing Chai, Jie Ma, Hong Liu. Deep Graph Attention Convolution Network for Point Cloud Semantic Segmentation[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210016 Copy Citation Text

show less

Fig. 1. Process of GAC. (a) Neighborhood points cluster toward center point; (b) calculation process of graph attention coefficient

Download full size

Process of dynamic graph convolution. (a) Center point looks for neighborhood point in three-dimensional space; (b) center point looks for neighborhood point in feature space

Fig. 2. Process of dynamic graph convolution. (a) Center point looks for neighborhood point in three-dimensional space; (b) center point looks for neighborhood point in feature space

Download full size

Fig. 3. Structure of DeepGAC

Download full size

Fig. 4. Qualitative results of semantic segmentation of S3DIS under different methods. (a) Original images; (b) ground truth; (c) ResGCN-28; (d) GAC; (e) ResGAC-28

Download full size

Type	PointNet	PointNet++	DGCNN	ResGCN-28	A-CNN	ResGAC-28 (ours)
OA	78.5	--	84.1	85.9	87.3	88.2
mIoU	47.6	53.2	56.1	60.0	62.9	64.5
Ceiling	88.0	90.2	--	93.1	92.4	92.6
Floor	88.7	91.7	--	95.3	96.4	96.5
Wall	69.3	73.1	--	78.2	79.2	81.1
Beam	42.4	42.7	--	33.9	59.5	36.8
Column	23.1	21.2	--	37.4	34.2	37.5
Window	47.5	49.7	--	56.1	56.3	59.3
Door	51.6	42.3	--	68.2	65.0	68.1
Chair	42.0	59.0	--	61.0	78.0	62.5
Table	54.1	62.7	--	64.9	66.5	69.1
Bookcase	38.2	45.8	--	51.5	56.9	54.7
Sofa	9.6	19.6	--	34.6	28.5	60.9
Board	29.4	48.2	--	51.1	48.0	66.7
Clutter	35.2	45.6	--	54.4	56.8	52.7

Table 1. Semantic segmentation results of 6-fold cross validation method on S3DIS unit: %

Type	PointNet	ResGCN-28	SPG^[21]	PointCNN^[22]	GAC	ResGAC-28 (ours)
OA	--	--	86.4	85.9	87.8	88.0
mIoU	41.1	52.5	58.0	57.3	62.9	63.3
Ceiling	88.8	--	89.4	92.3	92.3	92.4
Floor	97.3	--	96.9	98.2	98.3	97.5
Wall	69.8	--	78.1	79.4	81.9	80.1
Beam	0.1	--	0	0	0	0
Column	3.9	--	42.8	17.6	20.4	21.6
Window	46.3	--	48.9	22.8	59.1	60.4
Door	10.8	--	61.6	62.1	40.9	51.5
Chair	52.6	--	75.4	80.1	78.5	78.1
Table	58.9	--	84.7	74.4	85.8	85.9
Bookcase	40.3	--	69.8	66.7	61.7	64.6
Sofa	5.9	--	52.6	31.7	70.8	71.0
Board	26.4	--	2.1	62.1	74.7	69.8
Clutter	33.2	--	52.2	56.7	52.8	50.0

Table 2. Test results of different methods in region 5 of S3DIS unit: %

Aggregation mode	OA /%	mIoU /%
Max-pooling	86.0	60.8
Mean	83.3	55.1
ResGAC-28(ours)	88.2	64.5

Table 3. Test results for different aggregation modes

Formula mode	OA /%	mIoU /%
Spatial location information	87.9	62.8
Characteristic attribute information	87.7	62.6
ResGAC-28(ours)	88.2	64.5

Table 4. Test results of spatial location information and characteristic attribute information

Network structure	OA /%	mIoU /%
No Res-28	82.6	52.7
Fixed-28	86.1	60.8
ResGAC-28(ours)	88.2	64.5

Table 5. Validity test results of residual join and dynamic graph convolution

Parameter	mIoU /%	ΔmIoU /%	L	n	k
Reference	64.5	0	28	64	16
Depth	61.3	-3.2	7	64	16
Depth	62.2	-2.3	14	64	16
Width	58.6	-5.9	28	16	16
Width	61.5	-3.0	28	32	16
Neighbors	61.6	-2.9	28	64	4
Neighbors	62.4	-2.1	28	64	8

Table 6. Test results of model parameter adjustment

Yujing Chai, Jie Ma, Hong Liu. Deep Graph Attention Convolution Network for Point Cloud Semantic Segmentation[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210016

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information