Semantic Segmentation of LiDAR Point Cloud Based on CAFF-PointNet

Ming Lai; Jiankang Zhao; Chuanqi Liu; Chao Cui; Haihui Long

doi:10.3788/LOP202158.2028004

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 20 >Page 2028004 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 20, 2028004 (2021)

Semantic Segmentation of LiDAR Point Cloud Based on CAFF-PointNet

Ming Lai, Jiankang Zhao^*, Chuanqi Liu, Chao Cui, and Haihui Long

Author Affiliations

Department of Instrument Science & Engineering, School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

show less

DOI: 10.3788/LOP202158.2028004 Cite this Article Set citation alerts

Ming Lai, Jiankang Zhao, Chuanqi Liu, Chao Cui, Haihui Long. Semantic Segmentation of LiDAR Point Cloud Based on CAFF-PointNet[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2028004 Copy Citation Text

show less

Fig. 1. Overall framework of the model

Download full size

Fig. 2. Feature fusion module based on attention mechanism

Download full size

Fig. 3. Training and test sets (colored by height). (a) Training set; (b) test set

Download full size

Fig. 4. Loss function graph of training set and average accuracy graph of test set

Download full size

Fig. 5. Comparison of classification results between our model and ground truth. (a) Ground truth; (b) ours

Download full size

Fig. 6. Classification results of baseline and our model. (a) Ground truth ; (b) baseline; (c) ours

Download full size

Fig. 7. Comparison of classification results of different models. (a) Ground truth; (b) ours; (c) NANJ2; (d) RIT_1; (e) WhuY4

Download full size

Geometric features	L_λ	P_λ	S_λ	O_λ	E_λ	A_λ	Σ_λ	C_λ
Formula	$\frac{λ_{1} - λ_{2}}{λ_{1}}$	$\frac{λ_{2} - λ_{3}}{λ_{1}}$	$\frac{λ_{3}}{λ_{1}}$	$\sqrt[3]{λ_{1} λ_{2} λ_{3}}$	$\frac{λ_{1} - λ_{3}}{λ_{1}}$	- $\overset{3}{\sum_{i = 1}}$ λ_iln(λ_i)	$\frac{λ_{3}}{λ_{1} + λ_{2} + λ_{3}}$

Table 1. Calculation formulae for geometric features

Category	Training set	Test set
Powerline	546	600
Low vegetation	180850	98690
Impervious surfaces	193723	101986
Car	4614	3708
Fence/Hedge	12070	7422
Roof	152045	109048
Facade	27250	11224
Shrub	47605	24818
Tree	135173	54226
Total	753876	411722

Table 2. The number of points in each category for the training and test sets

Category	Powerline	Low vegetation	Impervious surfaces	Car	Fence	Roof	Facade	Shrub	Tree
Powerline	371	2	1	0	0	156	59	0	11
Low vegetation	2	86137	8247	198	301	463	205	2764	373
Impervious surfaces	4	9139	92287	76	49	335	15	71	10
Car	1	168	234	2945	138	138	22	62	0
Fence	1	944	222	67	2188	112	153	3457	278
Roof	60	1310	99	22	23	102102	609	578	4245
Facade	61	1156	136	142	16	2288	6193	733	499
Shrub	7	6571	255	199	717	812	444	12308	3505
Tree	47	2394	23	17	139	1093	619	5060	44834
Precision /%	66.9	89.8	90.9	80.3	61.2	94.9	74.4	49.1	83.4
Recall /%	61.8	87.2	90.4	79.4	29.4	93.6	55.1	49.5	82.6
F₁ score /%	64.3	83.4	90.7	79.9	39.8	94.3	63.4	49.4	83.0

Table 3. Confusion matrix of testing set classification results

Category	F₁ score
Category	Baseline	Ours(with GAM)	Ours(with AFF)	Ours(final)
Powerline	55.7	56.2	61.2	64.3
Low vegetation	80.7	80.8	82.1	83.4
Impervious surfaces	90.9	91.3	89.6	90.7
Car	77.8	73.6	75.2	79.9
Fence	30.5	28.4	34.5	39.8
Roof	92.5	94.5	93.8	94.3
Facade	56.9	55.3	60.3	63.4
Shrub	44.4	43.0	47.6	49.4
Tree	79.6	81.2	80.1	83.0
OA	82.2	82.6	83.2	84.8
Average of F₁	67.7	68.3	69.5	72.2

Table 4. Ablation experimental results of different models%

Category	F₁ score
Category	RIT_1^[24]	WhuY4^[25]	NANJ2^[26]	PointNet++^[18]	PointSIFT^[19]	Ours
Powerline	37.5	42.5	62	57.9	55.7	64.3
Low vegetation	77.9	82.7	88.8	79.6	80.7	83.4
Impervious surfaces	91.5	91.4	91.2	90.6	90.9	90.7
Car	73.4	74.7	66.7	66.1	77.8	79.9
Fence	18	53.7	40.7	31.5	30.5	39.8
Roof	94	94.3	93.6	91.6	92.5	94.3
Facade	49.3	53.1	42.6	54.3	56.9	63.4
Shrub	45.9	47.9	55.9	41.6	44.4	49.4
Tree	82.5	82.8	82.6	77	79.6	83
OA	81.6	84.9	85.2	81.2	82.2	84.8
Average of F₁	63.3	69.2	69.3	65.6	67.7	72.2

Table 5. Quantitative comparison between our model and other models%

Ming Lai, Jiankang Zhao, Chuanqi Liu, Chao Cui, Haihui Long. Semantic Segmentation of LiDAR Point Cloud Based on CAFF-PointNet[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2028004

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information