Semi-supervised Semantic Segmentation Based on UPS Strategy Self Training
To improve the effectiveness of semi-supervised semantic segmentation,this paper proposes a semi-supervised semantic segmentation network SPNS that combines loss normalization technology with UPS strategy.Using loss normalization techniques to alleviate the instability of self training in standard loss functions;the UPS strategy is a technique that combines uncertainty estimation and passive learning,by calculating the incompleteness of the output value as another threshold,reliable pseudo labels are selected,and finally the semi-supervised semantic segmentation task is completed using the generated pseudo labels and labeled data.The SPNS method has +2.06 improvement compared to training with only labeled data on the PASCAL·VOC dataset,and also has some improvement compared to other methods.
国家科技期刊平台
NSTL
万方数据
