Semi-supervised multi-organ image segmentation helps physicians enhance disease diagnosis and treatment planning, while reducing the effort required for organ annotation. However, the class-imbalance issue significantly limits the effectiveness of existing semi-supervised methods. In this study, we propose a two-stage co-training framework (TCF) that employs Cross Pseudo Supervision (CPS) strategy for self-pretraining and mean-teacher for post-training. Specifically, we introduce a multi-scale complementary feature dropout (MCFD) for 3D U-Net to improve feature learning for organs of varying sizes. Additionally, we develop 3D CutMix to implement strong augmentation in Stage 2 and present a high-confidence selecting (HCS) module to filter noise in pseudo labels, thereby enhancing the quality of pseudo labels for unlabeled data. Comprehensive experiments on two benchmark datasets, FLARE22 and AMOS22, demonstrate that TCF achieves superior performance compared to existing state-of-the-art methods. Code is available at https://github.com/stap1e/two-stage-for-semi-supervised-cpsandmeanteacher .