To improve the precision of the contour in instance segmentation, this study proposes an iterative contour deformation network (CD-Net) based on a graph convolutional network (GCN). The proposed method treats the segmentation results of the Mask R-CNN model as the initial contours and refines the instances contour iteratively. Specifically, a contour point set is first sampled from the initial contour. Considering the various sizes of the instances, and according to the size of corresponding bounding boxes determined by the Mask R-CNN, a local neighborhood graph is constructed for each selected contour point. Subsequently, multi-scales features are automatically selected and combined with features learned in Mask R-CNN for each point in the local neighborhood graph. The local neighborhood graphs with features are then fed into the GCN to learn the deformation vectors, and the instance contours are refined accordingly. Finally, the refined contour is treated as the initial contour, and the above process is repeated to obtain the final instance contours. The experimental results on the COCO and Cityscapes datasets demonstrate that the proposed method achieves state-of-the-art performance.

为了提高实例分割中轮廓的精度，本研究提出了一种基于图卷积网络（GCN）的迭代轮廓变形网络（CD-Net）。所提出的方法将Mask R-CNN模型的分割结果作为初始轮廓，并迭代地细化实例轮廓。具体来说，首先从初始轮廓中采样出轮廓点集。考虑实例的各种大小，根据Mask R-CNN确定的对应边界框的大小，为每个选定的轮廓点构造一个局部邻域图。随后，针对局部邻域图中的每个点，自动选择多尺度特征并与在 Mask R-CNN 中学习的特征相结合。然后将具有特征的局部邻域图输入 GCN 以学习变形向量，并相应地细化实例轮廓。最后将细化后的轮廓作为初始轮廓，重复上述过程，得到最终的实例轮廓。COCO 和 Cityscapes 数据集的实验结果表明，所提出的方法实现了最先进的性能。