Scene text detection is challenging due to the diverse text appearance, the complex background, and the expensive labeling of training data. For detecting arbitrary-shaped texts, most existing methods require heavy data labeling efforts to produce polygon-level annotations for supervised training. In order to reduce the cost in data labeling, we propose to combine center point annotation into mixed-supervised scene text detection, in which the dataset comprises small number of fully annotated images and large number of weakly annotated images by center points. For better incorporating point supervision, we adopt self-training strategy based on a detector which locates texts by predicting their centers. Besides, in order to weight the pseudo labels generated during self training, we also propose a novel regression uncertainty estimation module to measure the quality of detection results. Extensive experiments on five benchmark datasets (ICDAR2015, C-SVT, CTW1500, Total-Text and ICDAR-ArT) show that using small amount of polygon annotated data and large amount of center point annotated data, our detector can achieve competitive detection performance.