• Acta Photonica Sinica
  • Vol. 50, Issue 10, 1010003 (2021)
Saiqiang ZHANG1、2, Shaofeng SI1、2, Bin LU1、2, Qing LI2、*, Benyao CHEN3, and Anxin RONG3
Author Affiliations
  • 1Institute of Microelectronics of the Chinese Academy of Sciences,Beijing 100029,China
  • 2University of Chinese Academy of Sciences,Beijing 100049,China
  • 3Huzhou Special Equipment Inspection Center,Huzhou,Zhejiang 313000,China
  • show less
    DOI: 10.3788/gzxb20215010.1010003 Cite this Article
    Saiqiang ZHANG, Shaofeng SI, Bin LU, Qing LI, Benyao CHEN, Anxin RONG. A Person Detection Algorithm in Fisheye Images Based on Rotated Boxes[J]. Acta Photonica Sinica, 2021, 50(10): 1010003 Copy Citation Text show less
    Anchor-free person detection network under fisheye images
    Fig. 1. Anchor-free person detection network under fisheye images
    Different strategies used as bounding box regression samples
    Fig. 2. Different strategies used as bounding box regression samples
    Projection method for calculating rotated Gaussian kernel
    Fig. 3. Projection method for calculating rotated Gaussian kernel
    Comparison between fixed Gaussian kernel and adaptive Gaussian kernel
    Fig. 4. Comparison between fixed Gaussian kernel and adaptive Gaussian kernel
    Transformation from network regression outputs to predicted bounding boxes
    Fig. 5. Transformation from network regression outputs to predicted bounding boxes
    Minimum axis-aligned rectangle of two rotated bounding boxes
    Fig. 6. Minimum axis-aligned rectangle of two rotated bounding boxes
    Relationship between IoU and angle difference under different aspect ratios
    Fig. 7. Relationship between IoU and angle difference under different aspect ratios
    Comparison of qualitative results
    Fig. 8. Comparison of qualitative results
    Gaussian kernelMW⁃R/%HABBOF/%CEPDOF/%
    mAPAP75mAPAP75mAPAP75
    Axis-aligned Gaussian kernel31.923.340.535.217.16.4
    Rotated Gaussian kernel31.724.442.737.218.38.9
    Table 1. Comparative experiment of Gaussian kernel function
    Loss typeMW⁃R/%HABBOF/%CEPDOF/%
    mAPAP75mAPAP75mAPAP75
    PIoU43.633.949.247.927.814.8
    GIoU27.313.215.26.09.31.7
    CIoU42.331.749.546.826.214.8
    EIoU45.437.249.348.129.216.1
    AIoU46.440.350.049.429.617.0
    Table 2. Comparative experiment of IoU loss
    βMW⁃R/%
    mAPAP50AP75
    1.045.490.137.9
    1.345.290.336.3
    1.645.590.238.5
    1.946.490.940.3
    2.246.490.439.6
    2.545.890.637.3
    Table 3. Ablation experiment of hyperparameter
    Loss typeMW⁃R/%HABBOF/%CEPDOF/%
    mAPAP75mAPsmAPAP75mAPsmAPAP75mAPs
    PIoU43.633.934.349.247.954.727.814.818.3
    PIoU_A44.335.334.949.748.357.228.115.519.6
    CIoU42.331.733.049.546.853.426.214.815.9
    CIoU_A42.932.233.449.647.853.829.015.719.8
    EIoU45.437.233.449.348.153.729.216.119.0
    EIoU_A46.238.434.749.850.355.330.118.419.1
    AIoU46.440.336.750.049.455.829.617.019.7
    AIoU_A47.240.635.949.650.054.130.118.019.6
    Table 4. Comparative experiment of Gaussian kernel function
    MethodSizeFPSMW⁃R/%HABBOF/%CEPDOF/%
    mAPAP50AP75FmAPAP50AP75FmAPAP50AP75F
    RAPiD60820.553.496.850.994.156.796.762.695.838.283.127.679.3
    RAPiD1 02413.653.396.748.893.557.397.859.196.939.486.426.783.6
    RAPiD_2x60817.442.593.226.389.343.790.731.288.529.378.311.678.0
    AFRPD-3460878.553.395.953.093.854.195.956.793.236.684.223.781.8
    AFRPD-341 02439.055.597.955.996.259.497.263.594.440.186.129.284.5
    AFRPD-5360849.055.397.057.494.859.197.665.797.339.687.528.385.5
    AFRPD-531 02423.056.498.160.495.459.097.766.396.940.687.429.585.2
    Table 5. Comparison of the state-of-the-art algorithm and proposed algorithm
    Saiqiang ZHANG, Shaofeng SI, Bin LU, Qing LI, Benyao CHEN, Anxin RONG. A Person Detection Algorithm in Fisheye Images Based on Rotated Boxes[J]. Acta Photonica Sinica, 2021, 50(10): 1010003
    Download Citation