• Optics and Precision Engineering
  • Vol. 31, Issue 13, 1962 (2023)
Li CHEN1,2,3, Linhan LI1,2,3, Shiyong WANG1,3,*, Sili GAO1,3,*, and Xiangzhou YE1,2,3
Author Affiliations
  • 1Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai200083, China
  • 2University of Chinese Academy of Sciences, Beijing100049, China
  • 3Key Laboratory of Infrared System Detection and Imaging Technology, Chinese Academy of Sciences,Shanghai20008, China
  • show less
    DOI: 10.37188/OPE.20233113.1962 Cite this Article
    Li CHEN, Linhan LI, Shiyong WANG, Sili GAO, Xiangzhou YE. MMShip: medium resolution multispectral satellite imagery ship dataset[J]. Optics and Precision Engineering, 2023, 31(13): 1962 Copy Citation Text show less
    Processing flows of medium resolution multispectral satellite imagery ship dataset
    Fig. 1. Processing flows of medium resolution multispectral satellite imagery ship dataset
    Example of small scale data
    Fig. 2. Example of small scale data
    Example of large scale data
    Fig. 3. Example of large scale data
    Width and height statistics
    Fig. 4. Width and height statistics
    Area statistics
    Fig. 5. Area statistics
    Normalization methods
    Fig. 6. Normalization methods
    Experimental results
    Fig. 7. Experimental results
    DatasetsSourceNumberPixelClassResolution/mYear
    NWPUVHR-10Google Earth8001k×1k100.5-22014
    LEVIR16Google Earth22k800×60030.2-12017
    DOTAGoogle Earth/GF-2/JL-12 806800×800-4 000×4 00015≤0.52017
    NWPU-RESISC4517Google Earth31 500256×256450.2-302017
    HRRSD18Google Earth/Baidu Map21 761——130.15-1.22018
    DIORGoogle Earth23 463800×800200.5-302018
    xView19WorldView 31 4121 500×1 50060≤0.32018
    Table 1. Comparison of remote sensing data sets containing ships
    DatasetsSourceNumberPixelClassResolution/mBandYear
    HRSC2016Google Earth1 070300×300-1 500×900250.4-2RGB2016
    Kaggle20SPOT6/7,Pléiades192 556768×76810.3-6RGB2018
    MASATI21Microsoft Bing maps6 212512×5121——RGB2018
    FGSDGoogle Earth2 612930×930430.12~1.93RGB2020
    ShipRSImageNetMultiple3 435930×930-1 500×1 500500.12~6RGB2021
    FGSC-2322Multiple4 052——230.4~2RGB2021
    MMShipSentinel-25 016

    512×512

    10 980×10 980

    110RGB/NIR本文
    Table 2. Comparison of remote sensing ship data sets
    ConfigurationVersion
    Python3.7.11
    Pytorch1.8.0
    Torchvision0.9.0
    MMdetection2.22.0
    mmcv1.4.6
    CUDA11.1
    Table 3. Configuration information
    MethodBackboneNIR(%)RGB(%)Multispectral(%)
    AP50AP75AP90AP50AP75AP90AP50AP75AP90
    One-stage Anchor-based SSD512VGG1674.1144.921.4775.3941.673.0374.6244.164.89
    One-stage Anchor-based RetinaNetResNet5088.3246.953.0387.3938.902.2767.5041.863.41
    One-stage Anchor-free FoveaBoxResNet5082.5145.179.0976.9736.809.0979.1741.197.03
    One-stage Anchor-free FCOSResNet5089.1247.680.7188.9046.852.2788.9946.831.52
    Two-stage Faster R-CNNResNet5077.4150.882.6187.4649.400.7382.4152.772.14
    Two-stage Cascade R-CNNResNet5090.3953.311.3090.4652.723.0390.5752.139.09
    Table 4. Comparison of detection algorithms on test sets
    Methodlr
    SSD5120.000 5
    RetinaNet0.010 0
    FoveaBox0.005 0
    FCOS0.010 0
    Faster R-CNN0.020 0
    Cascade R-CNN0.020 0
    Table 5. Learning rate setting
    MethodBackboneRGB/%
    AP50AP50AP75
    Rotated-RetinaNetResNet5084.26359.0710.48
    Rotated-FCOSResNet5088.5265.767.93
    Rotated-Faster R-CNNResNet5078.9943.919.09
    Table 6. HRSC2016 test results
    Li CHEN, Linhan LI, Shiyong WANG, Sili GAO, Xiangzhou YE. MMShip: medium resolution multispectral satellite imagery ship dataset[J]. Optics and Precision Engineering, 2023, 31(13): 1962
    Download Citation