• Acta Optica Sinica
  • Vol. 40, Issue 21, 2115001 (2020)
Xuyi Zhang* and Jiale Cao
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
  • show less
    DOI: 10.3788/AOS202040.2115001 Cite this Article Set citation alerts
    Xuyi Zhang, Jiale Cao. Contour-Point Refined Mask Prediction for Single-Stage Instance Segmentation[J]. Acta Optica Sinica, 2020, 40(21): 2115001 Copy Citation Text show less
    PolarMask result visualization
    Fig. 1. PolarMask result visualization
    Network architecture of our overall framework
    Fig. 2. Network architecture of our overall framework
    Polar coordinates of the contour points
    Fig. 3. Polar coordinates of the contour points
    Semantic segmentation subnetwork
    Fig. 4. Semantic segmentation subnetwork
    Semantic segmentation of an image. (a) Original image; (b) semantic segmentation of area of concern
    Fig. 5. Semantic segmentation of an image. (a) Original image; (b) semantic segmentation of area of concern
    Network structures of different mask prediction subnetworks. (a) Using different prediction subnetworks (DPS); (b) using the same prediction subnetwork (SPS)
    Fig. 6. Network structures of different mask prediction subnetworks. (a) Using different prediction subnetworks (DPS); (b) using the same prediction subnetwork (SPS)
    Segmentation results of each stage. (a) Original image; (b) segmentation results of semantic segmentation subnetwork; (c) segmentation results of mask prediction subnetwork; (d) final instance segmentation results
    Fig. 7. Segmentation results of each stage. (a) Original image; (b) segmentation results of semantic segmentation subnetwork; (c) segmentation results of mask prediction subnetwork; (d) final instance segmentation results
    Instance segmentation of different methods. (a) Original image; (b) PolarMask; (c) our method
    Fig. 8. Instance segmentation of different methods. (a) Original image; (b) PolarMask; (c) our method
    Results of the proposed algorithm under the MS COCO test dataset
    Fig. 9. Results of the proposed algorithm under the MS COCO test dataset
    MethodAPAP50AP75APSAPMAPL
    Concat+1×1 conv30.551.031.912.532.746.3
    Sum30.851.532.312.432.045.2
    Table 1. Comparison of semantic segmentation feature fusion methodsunit: %
    MethodFigureAPAP50AP75APSAPMAPL
    DPSFig. 6(a)29.449.829.912.331.643.2
    SPSFig. 6(b)30.851.532.312.432.045.2
    Table 2. Comparison of experimental results of different mask prediction subnetworksunit: %
    λsegmλangleAPAP50AP75
    1.01.030.250.731.8
    0.51.030.050.531.8
    1.00.530.551.232.0
    1.00.330.651.332.2
    1.00.230.851.532.3
    1.00.130.751.332.3
    Table 3. Experimental results obtained by different loss function weightsunit: %
    MethodAPAP50AP75APSAPMAPL
    Baseline29.149.529.712.631.842.3
    Baseline+Semantic segmentation subnetwork29.150.630.111.730.644.7
    Baseline+Mask prediction subnetwork29.450.529.912.831.843.3
    Ours30.851.532.312.432.045.2
    Table 4. Comparison of each module under MS COCO-validation datasetunit: %
    MethodBackboneAPAP50AP75APSAPMAPL
    MNC[22]Resnet10124.644.324.84.725.943.6
    FCIS[5]Resnet10129.249.5-7.131.350.0
    YOLACT[12]Resnet10131.250.632.812.133.347.1
    PolarMask[16]Resnet10130.451.931.013.432.442.8
    OursResnet10132.553.634.313.134.348.0
    Table 5. Performance comparison of different methods under the MS COCO test datasetunit: %
    Xuyi Zhang, Jiale Cao. Contour-Point Refined Mask Prediction for Single-Stage Instance Segmentation[J]. Acta Optica Sinica, 2020, 40(21): 2115001
    Download Citation