LSNFS: A Local Feature Descriptor Algorithm with High Discrimination and Strong Robustness

Senda Hong; Haojie Cheng; Chunxiao Xu; Zhenxin Chen; Jiajun Wang; Lingxiao Zhao

doi:10.3788/LOP240666

Laser & Optoelectronics Progress, Volume. 61, Issue 22, 2215001(2024)

LSNFS: A Local Feature Descriptor Algorithm with High Discrimination and Strong Robustness

Senda Hong^1,2, Haojie Cheng^1,2, Chunxiao Xu^1,2, Zhenxin Chen², Jiajun Wang², and Lingxiao Zhao^2、*

¹School of Biomedical Engineering (Suzhou), Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei 230026, Anhui , China

²Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Science, Suzhou 215162, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Construction diagram of SNDS

Download full size

Fig. 2. Effectiveness evaluation of $A_{W N D A}$ attribute partitioning strategy. (a) Visual analysis of the objective function $s i n$ ; (b) RPC performance evaluation of LSNFS descriptor using different partitioning strategies on the U3OR dataset

Download full size

Fig. 3. Schematic diagrams of constructing LSNFS descriptor. (a) Key points and their local surface extraction; (b) LRA construction; (c) calculate 5 attribute values for each neighborhood point; (d) statistics on the distribution of attributes; (e) generation of sub histograms; (f) final histogram generation

Download full size

Fig. 4. Parameter settings of the LSNFS descriptor (the larger solid marker indicates the selected parameter values). (a) Test results on B3R; (b) test results on U3OR

Download full size

Fig. 5. Point cloud examples. (a) B3R dataset; (b) U3OR dataset; (c) U3M dataset; (d) QuLD dataset; (e) S3R dataset; (f) K3R dataset

Download full size

Fig. 6. RPC performance evaluation of eight descriptors on the B3R dataset. (a) 0.3 mr Gaussian noise; (b) 0.5 mr Gaussian noise; (c) 1/4 mesh decimation ; (d) 1/8 mesh decimation; (e) 0.3mr Gaussian noise+1/4 mesh decimation; (f) 0.5 mr Gaussian noise+1/8 mesh decimation

Download full size

Fig. 7. RPC performance evaluation of eight descriptors on U3M and U3OR datasets. (a) U3M dataset; (b) U3OR dataset

Download full size

Fig. 8. RPC performance evaluation of eight descriptors on S3R and K3R datasets. (a) S3R dataset; (b) K3R dataset

Download full size

Fig. 9. RPC performance evaluation of eight descriptors on QuLD dataset

Download full size

Fig. 10. Time required for feature extraction of eight descriptors under different support radii

Download full size

Fig. 11. Correct registration rate of 8 descriptors tested on U3M dataset

Download full size

Fig. 12. Four object registration cases based on LSNFS descriptor on U3M dataset. (a) Chef; (b) Chicken; (c) T-rex; (d) Parasaurolophus

Download full size

Fig. 13. Qualitative registration results based on LSNFS descriptor on 3DMatch, ETH, and TUM datasets. (a) (b) 3DMatch dataset; (c) (d) ETH dataset; (e) (f) TUM dataset

Download full size

Table 1. Descriptor parameter settings
View table
Table 1. Descriptor parameter settings
Condition $N_{r}$ $N_{h}$ $N_{W N D A}$ $N_{α}$ $N_{β}$
Setting $N_{r}$ 2-20 2 2 2 2
Setting $N_{h}$ 5 2-20 2 2 2
Setting $N_{W N D A}$ 5 5 2-20 2 2
Setting $N_{α}$ 5 5 10 2-20 2
Setting $N_{β}$ 5 5 10 13 2-20

Table 2. Characteristics of six standard public datasets

View table

Table 2. Characteristics of six standard public datasets

Dataset	Scanner	Challenge	Quality	Scenario	#Model	#Scene
B3R^［28-29］	Cyberware 3030 MS	Gaussian noise，mesh decimation	High	Retrieval	6	6
U3M^［30］	Minolta vivid 910	Occlusion	High	Registration		75
U3OR^［31］	Minolta vivid 910	Occlusion， clutter	High	Recognition	5	80
QuLD^［32］	NextEnigne	Qcclusion， clutter，real noise	Low	Recognition	5	80
S3R^［33］	SpaceTime Stereo	Occlusion， real noise， outliers	Low	Registration		57
K3R^［33］	Microsoft Kinect	Occlusion， real noise， outliers	Low	Registration		69

Table 3. Parameter settings for eight descriptors

View table

Table 3. Parameter settings for eight descriptors

Descriptor	Support radius /mr	Dimensionality	Length
TOLDI^［10］	20	3×20×20	1200
RoPS^［19］	20	3×3×3×5	135
MaSH^［21］	20	5×（10+8+10）	140
SDASS^［22］	20	5×5×15-30	345
SHOT^［27］	20	8×2×2×11	352
LFSH^［35］	20	15+10+5	30
RCS^［36］	20	6×12	72
LSNFS	20	5×5×10+5×13+5×15	390

Table 4. Feature matching recall rates of different methods on the 3DMatch dataset

View table

Table 4. Feature matching recall rates of different methods on the 3DMatch dataset

Method	Kitchen	Home1	Home2	Hotel1	Hotel2	Hotel3	Study	MIT Lab	Average
3DSmoothNet^［14］	97.0	95.5	89.4	96.5	93.3	98.2	94.5	93.5	94.7
SpinNet^［15］	99.2	98.1	96.1	99.6	97.1	100.0	95.6	94.8	97.6
MS-SVConv^［16］	99.6	99.4	94.2	99.1	99.0	100.0	95.9	100.0	98.4
CGF^［44］	60.3	71.1	56.7	57.1	53.8	83.3	37.7	45.5	58.2
PPF-Net^［45］	89.7	55.8	59.1	58.0	57.7	61.1	53.4	63.6	62.3
PPF-FoldNet^［46］	78.7	76.3	61.5	68.1	71.2	94.4	62.0	62.3	71.8
LSNFS	88.1	91.3	84.7	89.6	86.2	92.4	85.8	84.3	87.8

Tools

Get Citation

Copy Citation Text

Senda Hong, Haojie Cheng, Chunxiao Xu, Zhenxin Chen, Jiajun Wang, Lingxiao Zhao. LSNFS: A Local Feature Descriptor Algorithm with High Discrimination and Strong Robustness[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2215001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites