Fast pixel-matching for video object segmentation

Video object segmentation, aiming to segment the foreground objects given annotation of first frame, has been attracting increasing attentions. Many state-of-the-art approaches have achieved great performance by relying on online model updating or mask-propagation techniques. However, most models require high computational cost due fine-tuning during inference. Most based are faster but with relatively low failure adapt appearance variation. In this paper, we design a new make good balance between speed and performance. We propose model, called NPMCA-net, which directly localizes non-local technique matching pixels in reference target frames. Since bring information both previous frames, our network is robust large variation, can better occlusions. Extensive experiments show that approach achieve fast at same time (86.5% IoU DAVIS-2016 72.2% DAVIS-2017, 0.11s per frame) under level comparison. Source code available https://github.com/siyueyu/NPMCA-net.