基于孪生网络的快速视频目标分割

针对现有半监督视频目标分割方法不能同时满足分割精度和分割效率的问题，在传统半监督视频目标分割方法上引入注意力机制对分割结果进行修正. 首先，构建一个外观特征提取子网用于提取视频第1帧的特征图，并将其作为外观指导信息；然后，得到视频前一帧的分割结果，作为位置引导信息；最后，构建一个当前帧特征提取子网，以双分支的结构结合位置修正注意力与外观修正注意力，将位置信息和外观信息与当前帧特征图进行融合，实现目标分割. 实验结果表明，该目标分割方法可以纠正视频目标分割中的传播误差，并能有效提升分割精度.

Faster R-CNN: towards real-time object detection with region proposal networks

To solve the problem that the existing semi-supervised video target segmentation methods cannot ensure segmentation accuracy and efficiency at the same time, an attention mechanism into the general semi-supervised video target segmentation method was introduced to modify segmentation results. First, an appearance feature extraction subnet was constructed to extract feature map of the first frame of video and it was used as appearance guidance information. Second, the segmentation result of the previous frame was obtained and used as position guidance information. Finally, a current frame feature extraction subnet was constructed, which combined position correction attention and appearance correction attention in a double branch structure, so as to integrate the position information and appearance information into the current frame feature map and accomplish the target segmentation. Experiments show that the target segmentation method can correct the propagation errors in video target segmentation and improve the segmentation accuracy.