The label distribution in the real world often shows the long-tail effect, where a small number of categories account for the vast majority of samples. The temporal action detection problem is no exception. The existing temporal action detection methods often focus on the head categories with a large number of samples, while neglecting the few-sample categories. This study systematically defines the long-tail temporal action detection problem and proposes a weighted class-rebalancing self-training method (WCReST) based on a semi-supervised learning framework. WCReST makes full use of the large-scale unlabeled data that exists in the real world to rebalance the label distribution in the training samples to improve the model’s fit for the tail categories. Additionally, a pseudo-label loss weighting method is proposed for the temporal action detection task to enhance the stability of model training. Experiments are conducted on the THUMOS14 and HACS Segments datasets, using video samples from the THUMOS15 and ActivityNet1.3 datasets to form corresponding unlabeled datasets. In addition, the Dance dataset is collected to meet the application requirements of video review, which includes 35 action categories, 6632 labeled videos, and 13264 unlabeled videos, preserving the significant long-tail effect in data distribution. A variety of baseline models are used to conduct experiments on the THUMOS14, HACS Segments, and Dance datasets. The results demonstrate that the proposed WCReST can improve the model’s detection performance on tail action categories and can be applied to different baseline temporal action detection models to enhance their performance.