光学学报, 2020, 40 (21): 2115001, 网络出版: 2020-10-26
基于轮廓点掩模细化的单阶段实例分割网络 下载: 1076次
Contour-Point Refined Mask Prediction for Single-Stage Instance Segmentation
机器视觉 实例分割 语义分割 深度学习 卷积神经网络 角度预测 machine vision instance segmentation semantic segmentation deep learning convolutional neural network angle prediction
摘要
针对现有的实例分割方法PolarMask中分割结果边缘信息模糊的问题,通过对轮廓点角度偏置和距离的预测,基于轮廓点细化的单阶段实例分割网络准确提取出实例轮廓。同时,为了进一步提升实例分割的性能,利用语义分割子网络对实例边缘进行了进一步细化。实验结果表明,所提方法在大规模实例分割数据集MS COCO的测试集上的分割精度为32.5%,比现有的实例分割方法(PolarMask)提高了2.1个百分点,证明了所提方法的有效性。
Abstract
To solve the fuzzy problem of edge information in mask results by single-stage PolarMask, a contour-point refined network is proposed herein. By predicting the angel offset and distance for each contour point, a more accurate contour can be generated. Moreover, an extra semantic segmentation is added to further refine the edge information. Experiments show that the proposed method achieves a segmentation accuracy of 32.5% on the MS COCO test dataset, 2.1 percentages higher than the fundamental PolarMask, demonstrating the effectiveness of the proposed method.
张绪义, 曹家乐. 基于轮廓点掩模细化的单阶段实例分割网络[J]. 光学学报, 2020, 40(21): 2115001. Xuyi Zhang, Jiale Cao. Contour-Point Refined Mask Prediction for Single-Stage Instance Segmentation[J]. Acta Optica Sinica, 2020, 40(21): 2115001.