Researching | Learning Feature Point Descriptors for Detail Preservation

Laser & Optoelectronics Progress, Volume. 59, Issue 22, 2215002(2022)

Learning Feature Point Descriptors for Detail Preservation

Tao Long, Chang Su, and Jian Wang^*

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (10)

References (20)

Paper Information

Figures & Tables(10)

Fig. 1. Neural network architecture of feature point extraction

Download full size

Fig. 2. Schematic diagram of center offset of position prediction

Download full size

Fig. 3. Flow chart of homography estimation

Download full size

Fig. 4. Schematic diagram of homography error calculation

Download full size

Fig. 5. Failure case of baseline feature matching

Download full size

Fig. 6. Qualitative results of proposed method on images pairs on HPatches dataset. (a)Illumination cases.; (b) rotation cases; (c) perspective cases

Download full size

Table 1. Comparison of experimental results of different network structures
View table
Table 1. Comparison of experimental results of different network structures
Method Repeat LE HA-1 HA-3 HA-5 MS
Baseline 0.633 1.044 0.503 0.796 0.868 0.491
V1 0.675 0.831 0.505 0.822 0.897 0.576
V2 0.676 0.856 0.581 0.866 0.903 0.554
V3 0.669 0.842 0.586 0.871 0.912 0.555

Table 2. Comparison of key point detection performance of different methods

Table 2. Comparison of key point detection performance of different methods

Method	Repeatability rate		Localization error
Method	Low resolution	High resolution	Low resolution	High resolution
ORB	0.532	0.525	1.429	1.430
SURF	0.491	0.468	1.150	1.244
BRISK	0.566	0.505	1.077	1.207
SIFT	0.451	0.421	0.855	1.011
LF-Net（indoor）	0.486	0.467	1.341	1.385
LF-Net（outdoor）	0.538	0.523	1.084	1.183
SuperPoint	0.631	0.593	1.109	1.212
UnsuperPoint	0.645	0.612	0.832	0.991
Proposed method	0.669	0.663	0.842	0.926

Table 3. Comparison of homography estimation and matching performance of different methods

Table 3. Comparison of homography estimation and matching performance of different methods

Method	Low resolution，300 points				High resolution，1000 points
Method	HA-1	HA-3	HA-5	MS	HA-1	HA-3	HA-5	MS
ORB	0.131	0.422	0.540	0.218	0.286	0.607	0.71	0.204
SURF	0.397	0.702	0.762	0.255	0.421	0.745	0.812	0.230
BRISK	0.414	0.767	0.826	0.258	0.300	0.653	0.746	0.211
SIFT	0.622	0.845	0.878	0.304	0.602	0.833	0.876	0.265
LF-Net（indoor）	0.183	0.628	0.779	0.326	0.231	0.679	0.803	0.287
LF-Net（outdoor）	0.347	0.728	0.831	0.296	0.400	0.745	0.834	0.241
SuperPoint	0.491	0.833	0.893	0.318	0.509	0.834	0.900	0.281
UnsuperPoint	0.579	0.855	0.903	0.424	0.493	0.843	0.905	0.383
Proposed method	0.586	0.871	0.912	0.555	0.552	0.840	0.916	0.508

Table 4. Comparison of experimental results on different data subsets

Table 4. Comparison of experimental results on different data subsets

Method	Hpatches subset	Repeat	LE	HA-1	HA-3	HA-5	MS
Outlier_rejection^［14］	ALL	0.686	0.890	0.595	0.871	0.912	0.544
	Illumination	0.678	0.826	0.753	0.942	0.984	0.614
	Viewpoint	0.693	0.953	0.494	0.801	0.857	0.479
Proposed method	ALL	0.669	0.842	0.586	0.871	0.912	0.555
	Illumination	0.643	0.789	0.642	0.933	0.965	0.576
	Viewpoint	0.695	0.893	0.532	0.810	0.861	0.534

Tools

Get Citation

Copy Citation Text

Tao Long, Chang Su, Jian Wang. Learning Feature Point Descriptors for Detail Preservation[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2215002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Aug. 19, 2021

Accepted: Sep. 24, 2021

Published Online: Sep. 23, 2022

The Author Email: Jian Wang (jianwang@tju.edu.cn)

DOI:10.3788/LOP202259.2215002

Topics

laser devices and laser physics

Lasers and Laser Optics

laser manufacturing

Instrumentation, Measurement and Metrology