• Laser & Optoelectronics Progress
  • Vol. 59, Issue 22, 2215002 (2022)
Tao Long, Chang Su, and Jian Wang*
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
  • show less
    DOI: 10.3788/LOP202259.2215002 Cite this Article Set citation alerts
    Tao Long, Chang Su, Jian Wang. Learning Feature Point Descriptors for Detail Preservation[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2215002 Copy Citation Text show less
    Neural network architecture of feature point extraction
    Fig. 1. Neural network architecture of feature point extraction
    Schematic diagram of center offset of position prediction
    Fig. 2. Schematic diagram of center offset of position prediction
    Flow chart of homography estimation
    Fig. 3. Flow chart of homography estimation
    Schematic diagram of homography error calculation
    Fig. 4. Schematic diagram of homography error calculation
    Failure case of baseline feature matching
    Fig. 5. Failure case of baseline feature matching
    Qualitative results of proposed method on images pairs on HPatches dataset. (a)Illumination cases.; (b) rotation cases; (c) perspective cases
    Fig. 6. Qualitative results of proposed method on images pairs on HPatches dataset. (a)Illumination cases.; (b) rotation cases; (c) perspective cases
    MethodRepeatLEHA-1HA-3HA-5MS
    Baseline0.6331.0440.5030.7960.8680.491
    V10.6750.8310.5050.8220.8970.576
    V20.6760.8560.5810.8660.9030.554
    V30.6690.8420.5860.8710.9120.555
    Table 1. Comparison of experimental results of different network structures
    MethodRepeatability rateLocalization error
    Low resolutionHigh resolutionLow resolutionHigh resolution
    ORB0.5320.5251.4291.430
    SURF0.4910.4681.1501.244
    BRISK0.5660.5051.0771.207
    SIFT0.4510.4210.8551.011
    LF-Net(indoor)0.4860.4671.3411.385
    LF-Net(outdoor)0.5380.5231.0841.183
    SuperPoint0.6310.5931.1091.212
    UnsuperPoint0.6450.6120.8320.991
    Proposed method0.6690.6630.8420.926
    Table 2. Comparison of key point detection performance of different methods
    MethodLow resolution,300 pointsHigh resolution,1000 points
    HA-1HA-3HA-5MSHA-1HA-3HA-5MS
    ORB0.1310.4220.5400.2180.2860.6070.710.204
    SURF0.3970.7020.7620.2550.4210.7450.8120.230
    BRISK0.4140.7670.8260.2580.3000.6530.7460.211
    SIFT0.6220.8450.8780.3040.6020.8330.8760.265
    LF-Net(indoor)0.1830.6280.7790.3260.2310.6790.8030.287
    LF-Net(outdoor)0.3470.7280.8310.2960.4000.7450.8340.241
    SuperPoint0.4910.8330.8930.3180.5090.8340.9000.281
    UnsuperPoint0.5790.8550.9030.4240.4930.8430.9050.383
    Proposed method0.5860.8710.9120.5550.5520.8400.9160.508
    Table 3. Comparison of homography estimation and matching performance of different methods
    MethodHpatches subsetRepeatLEHA-1HA-3HA-5MS
    Outlier_rejection14ALL0.6860.8900.5950.8710.9120.544
    Illumination0.6780.8260.7530.9420.9840.614
    Viewpoint0.6930.9530.4940.8010.8570.479
    Proposed methodALL0.6690.8420.5860.8710.9120.555
    Illumination0.6430.7890.6420.9330.9650.576
    Viewpoint0.6950.8930.5320.8100.8610.534
    Table 4. Comparison of experimental results on different data subsets
    Tao Long, Chang Su, Jian Wang. Learning Feature Point Descriptors for Detail Preservation[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2215002
    Download Citation