Fig 1.
The HKMask framework for instance segmentation.
Fig 2.
The schema of the original residual module (a) and the hybrid kernel module (b). Hybrid kernel module introduces attention mechanism and mixed convolution on the basis of the original Resnet module.
Table 1.
Details of ResNet-50 and hybrid kernel module-50.
Fig 3.
The improved squeeze-excitation networks.
The max pooling is used to retain texture information and the average pooling retains global information of the feature map.
Fig 4.
Original remote sensing image (a) and corresponding binary image (b).
Fig 5.
Feature maps of three datasets in different stages.
Compared to the original method, ours has a significant suppressing effect on unrelated background pixels, while enhancing the pixels of the target instance. This is conducive to the convergence speed in the process of model training.
Table 2.
The gains of ISE component in our design.
Table 3.
The gains of hybrid kernel component in our design.
Table 4.
Quantitative results on COCO 2017val.
Fig 6.
Qualitative result of different methods on xBD and COCO datasets.